Methods and systems for predicting cell properties using imaging flow cytometer data and artificial intelligence

Abstract

Systems and methods for predicting and verifying properties of individual cells using label-free imaging and machine learning are disclosed. In an embodiment, a method for predicting and validating properties of individual cells includes acquiring image data of a mixture of cells and markers using an imaging flow cytometer, and identifying an individual cell by dispensing the mixture onto a cell placement or collection platform and matching a first marker sequence from the image data with a second marker sequence obtained from the platform. The method further includes inputting the image data into a machine-learning model configured to generate a predicted change in one or more properties of the identified cell, determining an actual change in the properties by measuring the cell at two different time points, and updating the machine-learning model by comparing the predicted and measured changes and incorporating the comparison result into the model.

Description

PCT ApplicationAttorney Docket No.: 009062.8568.WOOOMETHODS AND SYSTEMS FOR PREDICTING CELL PROPERTIES USING IMAGING FLOW CYTOMETER DATA AND ARTIFICIAL INTELLIGENCECROSS-REFERENCE TO RELATED APPLICATION

[0001] This patent document claims priority to and benefits of U.S. Provisional Patent Application No. 63 / 725,452, titled “PREDICTING CELL PROPERTIES WITH ARTIFICIAL INTELLIGENCE FROM IMAGING FLOW CYTOMETER DATA” and filed on November 26, 2024. The entire contents of the before-mentioned patent application are incorporated by reference as part of the disclosure of this document.STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0002] This invention was made with government support under GM 148128 awarded by the National Institutes of Health. The government has certain rights in the invention.TECHNICAL FIELD

[0003] The disclosed technology generally relates to cell analysis techniques.BACKGROUND

[0004] Predicting the properties of tissues or organisms from genomic data is a widely accepted practice in the medical and research communities. However, single-cell genomics falls short, as the RNA sequencing process destroys the cell, making it difficult to verify predictions on the same individual cell.SUMMARY

[0005] The disclosed technology provides systems and methods for predicting and verifying properties of individual cells using label-free imaging and machine learning.

[0006] In some embodiments of the disclosed technology, a method for predicting and validating properties of individual cells comprises: acquiring, by at least one processor, image data of a mixture of cells and markers using an imaging flow cytometer; identifying, by the at least one processor, an individual cell in the mixture on a cell placement or collection platform by dispensingPCT ApplicationAttorney Docket No.: 009062.8568.WOOO the mixture onto the cell placement or collection platform after acquiring the image data, and matching a first marker sequence derived from the image data with a second marker sequence obtained from the cell placement or collection platform; inputting, by the at least one processor, the image data into a machine learning model configured to generate a predicted change in one or more properties of the identified individual cell; determining, by the at least one processor, a change in the one or more properties of the identified individual cell on the cell placement or collection platform by measuring the one or more properties of the identified individual cell at a first time point and a second time point subsequent to the first time point and comparing measurements corresponding to the first and the second time points to identify the change in the one or more properties of the identified individual cell; and updating, by the at least one processor, the machine learning model by comparing the predicted change with the determined change to generate a comparison result and incorporating the comparison result into the machine learning model.

[0007] In some embodiments of the disclosed technology, a system for analyzing property changes of individual cells over time comprises: an imaging flow cytometer configured to capture an image of a mixture of cells and markers; a cell placement or collection platform configured to receive the mixture of cells and markers dispensed from the imaging flow cytometer; and at least one processor operably connected to the imaging flow cytometer and the cell placement or collection platform, the at least one processor configured to; acquire image data of the mixture of cells and markers from the captured image; identify an individual cell in the mixture on the cell placement or collection platform by matching a first marker sequence derived from the image data with a second marker sequence obtained from the cell placement or collectionplatform; input the image data into a machine learning model configured to generate a predicted change in one or more properties of the identified individual cell; determine a change in the one or more properties of the identified individual cell on the cell placement or collection platform by measuring the one or more properties of the identified individual cell at a first time point and a second time point subsequent to the first time point and comparing measurements corresponding to the first and the second time points to identify the change in the one or more properties of the identified individual cell; and update the machine learning model by comparing the predicted change with the determined change to generate a comparison result and incorporating the comparison result into the machine learning model.PCT ApplicationAttorney Docket No.: 009062.8568.WOOO

[0008] The above and other aspects and implementations of the disclosed technology are described in more detail in the drawings, the description, and the claims.BRIEF DESCRIPTION OF THE DRAWINGS

[0009] FIG. 1 illustrates an example cell analysis system, along with a methodology and workflow for predicting cell protein expression.

[0010] FIGS. 2A-2D illustrates an example of cell protein expression level prediction from day zero cell images and machine learning.

[0011] FIGS. 3A-3B illustrate an example workflow for predicting changes in cell properties following exposure to stress.

[0012] FIGS. 4A-4C illustrate detection of cell history and prediction of post-stress cell property variations.

[0013] FIGS. 5A-5D show results of a commercial flow cytometer for glucose deprived group in comparison with the control group at 0-hour and 72-hour after stress.

[0014] FIGS. 6A-6D show results of a commercial flow cytometer for thermal stressed group in comparison with the control group at 0-hour and 72-hour after stress.

[0015] FIG. 7 shows an example method 700 for predicting and validating properties of individual cells based on some embodiments of the disclosed technology.

[0016] FIG. 8 shows another example method 800 for predicting and validating properties of individual cells based on some embodiments of the disclosed technology.

[0017] FIG. 9 shows an example system for analyzing property changes of individual cells over time based on some embodiments of the disclosed technology.DETAILED DESCRIPTION

[0018] The disclosed technology provides systems and methods for predicting properties of individual cells using label-free imaging and machine learning, without requiring cell labeling. A three-dimensional imaging flow cytometer (3D-IFC) is employed to capture an image of a mixture of cells and markers (e.g., marker beads) to obtain the sequence of cells and markers.PCT ApplicationAttorney Docket No.: 009062.8568.WOOO

[0019] To enable observation of changes in individual cells over time, the system dispenses the imaged mixture of cells and marker beads, onto a position-tracked cell placement platform (CPP). Marker beads of distinguishable sizes or optical properties are used as spatial reference points. A sequence of marker beads and cells observed during imaging via the 3D-IFC is compared to the sequence observed on the CPP after dispensing. This matching allows precise mapping of each cell’s image data to its physical location on the platform. The mixture of cells and marker on the cell placement platform is then cultured under controlled conditions, and later re-imaged to measure outcomes such as protein expression level and response to stress.

[0020] These images are analyzed by a machine learning model, such as a fused convolutional autoencoder (CAE)-based classifier, to predict cell properties of interest, such as protein expression level or response to stress.

[0021] Unlike conventional methods that require destructive labeling or single-cell sequencing, the systems and methods implemented based on some embodiments of the disclosed technology enable label-free prediction and verification of individual cell properties by utilizing marker beads as positional references, thereby enabling the analysis of changes in a specific cell’s properties.

[0022] The use of marker bead-based spatial referencing eliminates the need for labeling each individual cell. This approach enables direct correlation between the initial label-free cell images and the experimentally observed outcomes after a period of incubation. The predicted cell properties are verified by comparing changes in fluorescence intensity or position over time, thus enabling accurate and scalable analysis of cell behaviors with minimal perturbation.

[0023] The disclosed technology provides methods, techniques, devices, systems and products that relate to an artificial intelligence (Al)-assisted prediction of the properties of individual cells. Some embodiments of the disclosed technology use Al to analyze single-cell images captured by a 3D imaging flow cytometer. Cell images are analyzed at day zero and Al is used to predict specific cell properties. These predictions are then evaluated as the cells continue to live and develop. Preliminary experimental results using some embodiments of the disclosed technology have demonstrated an accuracy of approximately 88% in predicting which individual cells will exhibit high levels of protein expression. The disclosed techniques may provide significant technical advantages in areas such as preventive medicine, drug development, cell therapy, andPCT ApplicationAttorney Docket No.: 009062.8568.WOOO biomedical research.

[0024] Some embodiments of the disclosed technology are designed to meet several important requirements, including minimally invasive analysis (e.g., favoring label-free detection), high- throughput data acquisition, high information content, and the ability to correlate analysis results with the physical locations of individual cells for subsequent verification.

[0025] In some embodiments, these requirements are addressed through a three-dimensional (3D) imaging flow cytometer, a robotically controlled cell placement platform, and a fused convolutional autoencoder-based classifier (CAE) neural network. These components enable experimental investigations into cell fate prediction.

[0026] A cell property prediction method based on some embodiments of the disclosed technology includes capturing and integrating information from various cellular features, such as morphology, signaling behavior, and gene expression, at different time points during cellular development to provide a comprehensive understanding of the regulatory mechanisms and molecular events that govern cell properties and behaviors.

[0027] In some implementations, supervised machine learning techniques are used, which rely on ground truth labels obtained from experimental observation. In other implementations, alternative learning models, such as unsupervised or self-supervised machine learning techniques, may be used to simplify network training and offers higher probabilities of discovering novel cellular states.

[0028] In some embodiments, the properties of each individual cell can be predicted. Singlecell genomics does not work in this context because the RNA sequencing process destroys the cell, making it impossible to verify our predictions. To test the hypothesis, the disclosed technology can be implemented in some embodiments to analyze single-cell images obtained from a 3D imaging flow cytometer using AL The cell image is analyzed at day zero and the Al-assisted cell property prediction is generated. The prediction is then examined later, as the cells continue to live and develop. In preliminary results, a method implemented based on some embodiments of the disclosed technology exhibited an accuracy of approximately 88% in predicting while cells would later exhibit high levels of protein expression. This technique can be applied in areas including preventive medicine, drug development, cell therapy, and fundamental biomedical research.PCT ApplicationAttorney Docket No.: 009062.8568.WOOO

[0029] In some embodiments, a cell analysis method includes: (1 ) sample preparation and data acquisition; (2) training of the Al model; and (3) prediction of the cell protein expression level.

[0030] In some embodiments, a cell analysis system comprises two interconnected hardware components: a 3D imaging flow cytometer (3D-IFC); and a robotic cell dispenser. The 3D-IFC can produce 3D side scattering and fluorescent images, as well as 2D transmission images of moving cells at a predetermined rate. When the cells exit the system, they are dispensed in a first- in-first-out manner on a template, together with marker beads that were added to the sample to help register cell locations. The maker bead sequences are used as references to relate the cell images to their corresponding cell locations on the cell placement membrane.

[0031] In some embodiments, cells that stably express fluorescent protein (e.g., green fluorescent protein, GFP) may be used. After imaging and dispensing them onto a position-tracked cell placement platform (CPP), the protein expression level may be measured by a fluorescence microscope at day zero (or 0 hour). The cells are then cultured on a membrane filter with a provision of full media in the incubator for a certain period of time (e.g., 48 hours). From these data, the relative protein production rate of each single cell over the certain period of time (e.g., the past 48 hours) is calculated. Based on the histogram of protein production rates, a simple gating is applied to separate the high expression level group (e.g., top 10%) from the rest of the population.

[0032] In an embodiment, a certain number of cell images are used to train the Al model or neural network to classify cells into a high protein production rate group and a low or ordinary protein production rate group using supervised learning. In an implementation, the images used for training may consist of unlabeled 3D side scattering channel (SSC) and 2D transmission images taken from the 3D-intelligent flow cytometry (IFC) at Ohr (at day zero) before the 48-hour culture. Through the training process, the neural network encodes the input images to a latent convolutional layer. During inference, the network uses the flattened latent channels to make predictions.

[0033] Prediction of the properties of tissues or organisms from genomics data is widely accepted by the medical field. Some embodiments of the disclosed technology address a different question: whether the properties of each individual cell can be predicted. Conventional single-cell genomics is not suitable for this purpose because the RNA sequencing process destroys the cell, thereby preventing verification of any prediction. To evaluate this hypothesis, some embodiments of the disclosed technology investigate an approach that utilizes artificial intelligence (Al) toPCT ApplicationAttorney Docket No.: 009062.8568.WOOO analyze single-cell images obtained from a 3D imaging flow cytometer. The cell image is analyzed at day zero, and an Al-assisted cell property prediction is generated. The prediction is then examined as the cells continue to live and develop. Preliminary results indicate an accuracy of 88% in predicting cells that will exhibit a high protein expression level. This technique can have significant ramifications for preventive medicine, drug development, cell therapy, and fundamental biomedical research.

[0034] Understanding cellular behavior at the single-cell level serves as a fundamental basis for advancements in modern cell biology, biotechnology, and therapeutic development. To date, research has been focused on studies of the molecular or morphological properties of cells at the time of analysis, limited by the methods of single-cell genomics or proteomics that destroy or seriously disturb the cells. While such analyses provide valuable information, existing technologies lack the capability to predict the cell behaviors at the single-cell level.

[0035] An important question is considered here: what if one can analyze the cell at one time and then predict the cell behaviors in the future? If the fate of each individual cell can be accurately predicted, a wide range of applications becomes possible. For example, in the manufacturing of monoclonal antibody drugs, the current technique requires the selection of single cell derived micro colonies having a high proliferating rate and protein drug production rate. Presently, a single cell must be placed in each well in well plates for tens of thousands of times, followed by days or weeks of culturing and measurements of the cell properties, before finally selecting those singlecell derived micro colonies exhibiting high performance for drug production. If the proliferation rate and protein production rate of each cell could be predicted at day zero, this could result in significant reductions in cost and labor associated with drug manufacturing. Cell property prediction capabilities can also have impact in the field of preventive medicine and personalized medicine. If one can predict the likelihood that cells in a sub-healthy or precancerous state may become cancerous, preventive treatments can be administered to lower the risks. In such cases, cell property prediction has the potential not only to provide biological insights but also to enable actionable plans to benefit health and prevent diseases. The single-cell property prediction methodology implemented based on some embodiments of the disclosed technology complementary to genetic screening techniques. While genetic screening operates at the tissue and organ levels, some embodiments of the disclosed technology enable cell property prediction at the single-cell level. When integrated with single-cell RNA sequencing and single-cellPCT ApplicationAttorney Docket No.: 009062.8568.WOOO proteomic analyses, this technique can significantly advance the understanding of cellular biology.

[0036] Establishing the cell property predication capabilities at the single-cell level requires a unique combination of technologies. Unlike most cell analysis methods that produce information at the endpoint, technologies designed for cell property prediction must preserve cell integrity after analysis to allow subsequent verification and assessment of the predictions. To minimally affect cell properties, label-free single-cell imaging appears to be a suitable technique. It is also important to isolate each imaged cell and place it in an environment suitable for the cell to continue its development. The above requirements suggest that an imaging flow cytometer capable of producing label-free 3D images of individual cells at high throughput is a desirable tool for this purpose. Unlabeled or minimally labeled cells experience minimum unintended cell perturbations. The flow cytometer platform supports single-cell imaging, and the resulting 3D images provide rich information that serves as a basis for cell property prediction. In some embodiments, artificial intelligence (Al) may form the foundation of cell property prediction. In some embodiments, a 3D imaging flow cytometer is used in combination with a position-tracked cell placement platform (CPP) as the hardware, and a fused convolutional autoencoder-based classifier (CAE) architecture to perform cell property prediction.

[0037] In some embodiments, a flow cytometer system configured to generate 3D single-cell images based on a scanning light sheet and an array of pinholes configured in a confocal microscope design. The system combines the high-throughput capabilities of flow cytometry and the imaging capabilities of microscopy. The setup supports 3D side scattering and fluorescent images, as well as 2D transmission images. As a label-free method, the 3D side scattering image is generated by sectioning a cell into 10 slices and imaging the scattering centers on each of the slices, thus creating more information contents than 2D images. The data is then processed and interpreted by a machine learning model that deciphers complex patterns within the data and generates predictions of the cell properties. In addition, a cell dispensing platform that can track the cell image and the cell location for each single cell enables observation of cellular behaviors over time and evaluation of the prediction accuracy. To validate the concept, two experiments, (1) predictions of protein expression level and (2) cell property changes after stresses, were performed. The latter also includes the ability to detect a cell’s history (e.g., whether a cell has experienced stress in the past).PCT ApplicationAttorney Docket No.: 009062.8568.WOOO

[0038] In the first experiment, prediction of protein expression levels was performed on individual Chinese Hamster Ovary (CHO) cells. From the 3D scattering and 2D transmission images of green fluorescent protein (GFP) transfected CHO cells obtained immediately after harvesting from the culture medium (e.g., day zero), without measuring the green fluorescent protein (GFP) expression level at the time of harvest, the system was able to correctly identify, in most cases, those CHO cells that will later produce more GFPs than the rest of the population. The high GFP expression level group (e.g., top 10%) was successfully distinguished from the low and ordinary GFP expression level groups at a balanced prediction accuracy of 88.2%.

[0039] In the second set of experiments, the cells were divided into two groups: cells that had been subjected to stress; and cells that had not. The experiment was performed to determine whether a given cell had been subjected to stress (e.g., to infer the cell’s history) and to predict how the stressed cells would behave differently in the future. A blind test was conducted in which the testers did not know whether the samples had been stressed or not. As a result, the testers were required to predict not only whether the properties of each cell would change, but also whether the cells imaged at day zero had a history of exposure to stress. Two types of stressors were applied to CHO cells: elevated temperature (40°C) and glucose deprivation. For the thermal-stressed group, a prediction balanced accuracy of 81.8% was achieved. For the glucose deprivation group, a balanced prediction accuracy of 99.4% was achieved. To evaluate the prediction accuracy, each group of cells was harvested 72 hours after the prediction and analyzed using a conventional flow cytometer.

[0040] Prediction of cell protein expression level

[0041] In some embodiments, the experiment comprises three steps: sample preparation and data acquisition; training of the Al model; and prediction of cell protein expression level.

[0042] In some embodiments, a cell analysis system comprises a 3D imaging flow cytometer and a robotic cell dispenser that are coupled to one another. The 3D-IFC system can produce 3D side scattering and fluorescent images as well as 2D transmission images of moving cells at a rate of 1000 cells / s. When the cells exit the system, they are dispensed in a first- in-first-out manner on a template together with the marker beads added to the sample to help register the cell location.

[0043] FIG. 1 illustrates an example cell analysis system, along with a methodology and workflow for predicting cell protein expression. Specifically, FIG. 1 shows the overall schematicPCT ApplicationAttorney Docket No.: 009062.8568.WOOO of a cell analysis system 110, and maker bead sequences 120 that are used as references to relate the cell images to the cell locations on the cell placement membrane.

[0044] Referring to FIG. 1, CHO cells that stably express humanized enhanced green fluorescent protein (eGFP) is used. After imaging and dispensing them on the CPP. the GFP protein expression level is measured by a fluorescence microscope at day zero (or 0 hour). The cells were cultured on the membrane filter with a provision of full media in the incubator (with 37°C, 5% CO2) for 48 hours. FIG. 1 also shows the sequence extraction results 120 from both the image flow cytometer (IFC) and cell placement platform (CPP) and the GFP FL protein expression level at 0 hour and 48 hours. An example of cells and beads on the CPP at 0 hour (132) and 48 hours (134) are shown in FIG. 1.

[0045] Cells and marker beads mixtures are examined by the 3D imaging flow cytometer (3D- IFC) (110). In FIG. 1, AOD indicates acousto-optic deflector, CL indicates cylindrical lens, IO indicates 20x / 0.42 illumination objective. SDO indicates 10x / 0.28 side detection objective, SSP indicates side spatial filter, DMs indicate dichroic mirrors, FDO indicates forward detection objective, FSP indicates forward spatial filter, DIG indicates 125 MS si digitizer, and PMT indicates photomultiplier tube. Samples are then dispensed on the cell placement platform (CPP) (130) in a first- in-first-out manner after exiting the 3D-IFC (110). The order of cells and marker beads of different sizes or colors (represented here by A, T, C) from the 3D-IFC signals and CPP are compared for the purpose of matching the cell images to the cell positions and eliminating any sections showing deletion or displacement errors. An example of matched sequences is shown in FIG. 1. A numerical number is used to represent the number of cells between two marker beads. For example, a sequence C / 2 / T represents that 2 cells exist between a marker bead C and a marker bead T. The sequence from the IFC is identical to the sequence from the CPP if there are no deletion (e.g., missing cells) or displacement (e.g. cell position scrambling) errors. Any error sections will not be included in the analysis. The table below the IFC and CPP sequences 120 shows the protein expression level of each single cell at 0 hr (122) and 48hr (124), and the difference in the value shows the protein production rate (126). FIG. 1 also shows schematic illustrations (132, 134) of what occur to each single cell on the CPP over time. Each cell is developed into a single-cell derived microcolony and produces more GFPs over time. Cells on the CPP are supplied with full growth media in an incubator at 37°C with 5% CO2.PCT ApplicationAttorney Docket No.: 009062.8568.WOOO

[0046] From these data, the relative GFP protein production rate of each single cell over the past 48 hours is calculated. Based on the histogram of protein production rate, a simple gating is applied to separate the high (top 10%) expression level group from the rest of the population.

[0047] FIGS. 2A-2D illustrates an example of CHO cell protein expression level prediction from day zero cell images and machine learning. Specifically, FIG. 2A shows, from the histogram 210 of GFP protein expression rate, the high production group and low production group are gated out. Representative 3D SSC images 212 from each group are shown. FIG. 2B shows the machine learning model used to analyze and classify the 3D SSC images. Through the training process, the model learns to classify high GFP production cells and low / ordinary GFP production cells using the 3D SSC images captured at 0-hour. FIG. 2C shows an example of visualization of the latent encoded features from the trained convolutional autoencoder. FIG. 2D shows an example of confusion matrix of the machine learning model for performance evaluation. A balanced accuracy of 88.2% was achieved in prediction of single cell protein production capability.

[0048] FIG. 2A shows the histogram of the protein expression from single cells and some representative 3D SSC images in each group. In some embodiments, 3D SSC images are used as inputs for a machine learning model. In an example, the machine learning model includes a convolutional autoencoder-based classifier (CAE) to train the neural network. The results provide the ground truth of the GFP generation rate of each CHO cell.

[0049] In some implementations, a total of 1000 cell images are used to train the ML model to classify the CHO cells into the high GFP production group and the low / ordinary GFP production group using supervised learning. The images used for training are unlabeled 3D SSC and 2D transmission images acquired from the 3D-IFC at Ohr (e.g., day zero) before the 48-hour culture.

[0050] Through the training process, the neural network encodes the input images into a latent convolutional layer. During inference, the network utilizes the flattened latent channels to generate predictions. FIGS. 2C and 2D show the t-SNE visualization result and the corresponding confusion matrix of the inferencing results that predict which group each GFP-transfected CHO belongs to. The classification report shows a prediction accuracy rate of 88.2%. In this way, the disclosed technology can be implemented in some embodiments to predict the protein production rate of individual cells without labeling, enabled by a unique combination of hardware and machine learning.PCT ApplicationAttorney Docket No.: 009062.8568.WOOO

[0051] Prediction of cell property variations after stress

[0052] Predicting changes in cell properties based on cellular history can provide new insight for disease development, which can be valuable for preventive medicine. In the fields of drug response and cell therapy, predicting cellular properties days or weeks before the responses actually happen can enable early selection of target cells and contribute to the design of more effective treatments.

[0053] In an experiment conducted to investigate this concept, CHO cells were subjected to two types of stress: thermal stress and nutrient deprivation. These stresses were applied for a relatively short time period so as not to cause any obvious effects on the cell’s health or viability. Label-free images of the stressed and unstressed cells are acquired before any apparent phenotypical changes occur, and the images captured at day zero are used to predict whether the analyzed cells would eventually exhibit effects associated with the applied stress.

[0054] FIGS. 3A-3B illustrate an example workflow for predicting changes in cell properties following exposure to stress. Specifically, referring to FIG. 3A, cell culturing and stress introduction are performed. Both stressed and control samples are examined by a flow cytometer 310. Referring to FIG. 3B, after 6 hours of stress treatments, cells are harvested and examined using a 3D imaging flow cytometer (3D-IFC) 320 (e.g., day zero) and the resulting 3D side scattering channel (SSC) images 330 are used to predict the cell behaviors using a machine learning (ML) model. Example single cell images from stressed and control groups are shown. Each 3D cell image is divided into 10 cross-sections showing the scattering patterns of each crosssection.

[0055] FIGS. 4A-4C illustrate detection of cell history and prediction of post-stress cell property variations. Specifically. FIG. 4A illustrates a fused convolutional autoencoder (CAE) based classifier 410 in which two CAE structures 412, 414 are combined, each receiving a single modality image as input. In the latent space, the latent features are concatenated and used to make predictions. FIG. 4B is a visualization of the latent encoded features 420 from the trained fused CAE-based classifier 412 for the thermal-stressed group. The ability to detect cell history (e.g., whether a cell has experienced thermal stress or not) is summarized in the confusion matrix 430, showing a balanced accuracy of 81.6%. FIG. 4C is a visualization of the latent encoded features 422 from the trained fused CAE-based classifier 414 for the glucose-deprived group. The abilityPCT ApplicationAttorney Docket No.: 009062.8568.WOOO to detect cell history (e.g., whether a cell has experienced glucose deprivation stress or not) is summarized in the confusion matrix 432, showing a balanced accuracy of 99.4%.

[0056] Referring to FIGS. 3A-3B, the overall workflow of the experiment is illustrated. Fortyeight hours prior to the stress treatment, 3 dishes of CHO cells are seeded with an initial concentration of 105 / mL cells. After 48 hours of culturing in a 37°C, 5% CO2 incubator supplemented with full culture media, cells are in the logarithmic growth phase. One dish of cells is transferred to an incubator in 40°C, 5% CO2 to thermally stress the cells. The other dish of cells is supplied with the culture media without glucose. Another dish is used as a control. Both the control dish and the glucose-deprived dish are placed in a 37 °C, 5% CO2 incubator. After 6 hours, cells from all three dishes are harvested and examined by the 3D IFC to generate “day zero” images. FIG. 3B shows examples of 3D SSC single cell images having each cell image divided into 10 cross sections. Subsequently, all cells are cultured in a 37°C, 5% CO2 incubator in full culture media. Each group of cells was then examined using forward scattering (FSC) and back scattering (BSC) signals of a flow cytometer (e.g., WOLF system, NanoCellect Biomedical Inc.) at a 24-hour interval and up to 72 hours to monitor any morphological changes for stressed cells

[0057] FIGS. 5A-5D show results of a commercial flow cytometer for glucose deprived group in comparison with the control group at 0-hour and 72-hour after stress. Specifically, FIGS. SA- 50 show back scattering (FIGS. 5A-5B) and forward scattering (FIGS. 5C-5D) signals from glucose deprived cells at Ohr and 72 hours. A modest shift of signal distribution for glucose deprived cells appears 72 hours after stress.

[0058] FIGS. 6A-6D show results of a commercial flow cytometer for thermal stressed group in comparison with the control group at 0-hour and 72-hour after stress. Specifically. FIGS. 6A- 6D show back scattering (FIGS. 6A-6B) and forward scattering (FIGS. 5C-5D) signals from thermal-stressed cells at Ohr and 72 hours. 72-hour. The signal shift for thermal-stressed cells is not obvious. The results in FIGS. 5A-5D and FIGS. 6A-6D indicate that commercial How cytometers are unable to detect cell morphological changes after stress without labeling.

[0059] FIGS. 5A-5D and FIGS. 6A-6D show comparisons of FSC and BSC signals from a commercial flow cytometer between the control group and the glucose-deprived and thermal- stressed groups at 0 hour and 72 hours. At 0 hour, immediately following the application of stress, minimal differences among the cell groups are detectable using a commercial flow cytometer.PCT ApplicationAttorney Docket No.: 009062.8568.WOOOAfter 72 hours, a modest shift in the histograms of FSC and BSC signals is observed in the glucose- deprived cells, while the thermal-stressed cells exhibit even smaller changes in signal distribution relative to the control group. These results indicate that even with continuous monitoring of cell properties for 72 hours, it remains difficult to determine whether an individual cell has experienced stress or undergone stress-induced morphological changes using a commercial flow cytometer without the aid of specific biomarker labeling. The disclosed technology can be implemented in some embodiments to investigate whether it is possible to detect a cell’s history (e.g., whether it has been stressed or not) and predict its future morphological changes using 3D label-free cell images and machine learning.

[0060] In some embodiments, a fused convolutional autoencoder, in which 2D CAE and 3D CAE are combined, each configured to receive a single image modality as the input, is used to detect cell (stress) history and predict post-stress cell properties. The convolutional layers extract image features, while the max-pooling kernels encode them into the subsequent layers during the contracting paths. The latent spaces from the two CAEs are combined by concatenation, followed by a fully connected layer and a Softmax layer to make a classification decision. FIG. 4A shows an example of the network structure.

[0061] Experiments were conducted to detect whether a cell was stressed and predict its future morphological changes for both glucose deprived and thermally stressed groups using the cell images taken at “day zero” and the fused CAE model. The results of the thermally stressed and glucose-deprived groups are shown in FIGS. 4B and 4C, respectively. The confusion matrices show that the fused CAE model was able to accurately detect the history of the cells and predict cell property changes for the stressed cells at day zero. The fused CAE demonstrates an averaged balanced prediction F_1 score of 0.994 for the glucose-deprived cells across all the Cross Validation (CV) folds. The prediction of cell property changes for thermally stressed cells gives rise to an averaged balanced prediction F_1 score of 0.814 for all CV folds.

[0062] Methods and Materials

[0063] Preparation of CHO cells for protein expression prediction

[0064] CHO cell line that stably expressed the humanized enhanced green fluorescent protein (eGFP) was used for the protein expression level prediction experiment. Cell lines were cultured with growth media (e.g., Ham’s F12K (Kaighn’s) Medium, 10% Fetal Bovine Serum, 1%PCT ApplicationAttorney Docket No.: 009062.8568.WOOOAntibiotic-antimycotic) in a 10 cm petri dish to 90% confluency before harvesting. After culturing, cells were harvested and resuspended to a concentration of approximately 5xl05cells / mL in IX PBS. Prior to introduction into the 3D-IFC. cells were mixed with marker beads (e.g., lOum, 20um, 30um) at a cell-to-bead ratio of around 2:1. After dispensing the cells and the marker beads onto the cell placement platform (CPP), cells were cultured on the membrane filter in growth media. GFP protein expression level of each cell was registered at 0 hour and 48 hours using a fluorescence microscope.

[0065] Preparation of CHO cells for detection of cell stress and prediction of post-stress cell property changes

[0066] Depending on cell concentration, the growth curve of CHO cells normally consists of a logarithmic cell proliferation phase, a stationary phase, and a death phase. Cells normally enter the exponential cell proliferation phase 24 to 72 hours after seeding. In this experiment, 48 hours prior to the stress treatments, 3 dishes of CHO cells were seeded with an initial concentration of 105 / mL cells. After 48 hours of culturing in a 37 °C, 5% CO2 incubator supplemented with full growth media (e.g., DMEM with 25mM D-Glucose, 10% Fetal Bovine Serum, 1% Antibiotic- antimycotic), cells entered the logarithmic growth phase. One dish of cells was transferred to an incubator in 40°C, 5% CO2 for thermal stress. Another dish of cells was supplemented with the culture media without glucose (DMEM no Glucose, 10% Fetal Bovine Serum, 1% Antibiotic- antimycotic). The other dish was used for the control sample. The glucose deprived dish and the control dish were cultured in a 37 °C, 5% CO2 incubator. After 6 hours of incubation for thermal stress and glucose deprivation for two of the three dishes, cells were harvested and resuspended to a concentration of approximately 5xl05cells / mL in IX PBS.

[0067] Customized CAE-based classifier for image classifications

[0068] A deep autoencoder (DAE) includes an encoder and a decoder, each including two fully connected layers stacked together. In contrast, a deep convolutional autoencoder (DCAE) incorporate convolutional and deconvolutional layers into both the encoder and decoder, instead of solely relying on fully connected layers. This approach is designed to leverage the benefits of a convolutional neural network for image processing, particularly for 3D image processing. Convolutional neural networks exhibit distinguished properties such as sparse interactions, parameter sharing, and equivariant representations, which are crucial for effectively translatingPCT ApplicationAttorney Docket No.: 009062.8568.WOOO latent features.

[0069] In some embodiments, two CAE-based classifiers, the 3D CAE-based classifier and the fused CAE-based classifier, are developed. The fused CAE-based classifier combines two CAE structures (2D CAE and 3D CAE), each configured to receive a single image modality as the input. The convolutional layers extract image features, while the max-pooling kernels encode them into the subsequent layers during the contracting paths. The latent spaces from the two CAEs are combined by concatenation and followed by a fully connected layer and a Softmax layer to enable a classification decision. The reconstructed images output from the decoders are leveraged to optimize the loss function during the model training process. The architecture of the fused CAEbased classifier is illustrated in FIG. 4A. The 3D CAE-based classifier can be considered as part of the fused network, while the concatenation in the latent space is eliminated. The output of the Softmax layer can be written as:where x is the input vectors, C is the number of classes.

[0070] For the first 20 epochs of training, the CAE-based classifier was trained to minimize the mini-batch averaged mean-square error (MSE) loss between the input images and generated output images. The fused CAE-based classifier was optimized using a weighted MSE loss including the contributions of 2D and 3D CAE. The mini-batch averaged mean-square error loss can be expressed as:where x2dand xdare the input 2D transmission image and model-generated image, respectively. M2dis the flattened transmission image vector dimension, x3dand x3dare the input 3D SSC image and model-generated image, respectively. M3dis the flattened SSC image vector dimension. N is the batch size in the mini-batch. w1= 0 is the mean-square-error loss weight assigned to the two image modalities. In the 3D-CAE based classifier, w_l=0 since there is no 2D image input.PCT ApplicationAttorney Docket No.: 009062.8568.WOOO

[0071] As another component in the overall loss function, the averaged cross-entropy loss can be written asis the ground truth class vector, y, is the predicted class vector, and N is the data size in the mini-batch.

[0072] The final weighted loss function, defined in (4), was adopted to train the network.L = w2■ LCE+ (1 — w2) ■ LMSE(4) where w2is the weight coefficient for the cross-entropy term relative to the MSE loss.

[0073] In some embodiments, the machine learning models may be implemented using deep learning frameworks such as PyTorch and trained using computing hardware, including commercially available CPUs and GPUs with sufficient memory and processing capabilities.

[0074] In some embodiments, a novel approach for predicting properties of individual cells. To ensure such predictive capability, the disclosed technology can be implemented in some embodiments to satisfy the following requirements: analysis with minimum disturbances to the cells (e.g., favoring label-free detection), high throughput, high information content, and an ability to correlate the cell analysis data with the cell location for assessment of the prediction results. Some embodiments of the disclosed technology can provide a combined system that includes a 3D imaging flow cytometer, a robotically controlled cell placement platform, and a fused CAE neural network to meet the above requirements. The combined system can enable the experimental study of cell fate prediction. As the technology advances and more studies in this area are conducted, the disclosed technology may contribute to a deeper understanding of cell biology. Furthermore, the predictive capabilities may find practical applications in areas such as pharmaceutical manufacturing and preventive medicine. In the longer term, the ability to predict cell fate may also contribute to the development of cell therapy and regenerative medicine.

[0075] The methods and systems for predicting cell properties based on some embodiments can capture and integrate information from multiple cellular features, such as morphology, signaling pathways, and gene expression, at different time points during cellular development.PCT ApplicationAttorney Docket No.: 009062.8568.WOOOThis integration provides a comprehensive understanding of the regulatory mechanisms and molecular events that govern cell properties and behaviors. The current approach uses supervised ML that relies on the ground truth labels. The disclosed technology can also be implemented in some embodiment to use unsupervised learning or self-supervised learning, which simplifies the network training and offers higher probabilities of discovering novel cellular states.

[0076] As explained above, the experimental results presented in this patent document demonstrate the feasibility of predicting cell properties at the single-cell level. Notable results include a prediction accuracy of approximately 88% for identifying high protein-producing cells, as well as balanced prediction accuracies of approximately 82% and 99% for detecting cell stress history and post-stress cell properties pave, respectively. These results establish a foundation for future developments in this emerging area of research.

[0077] FIG. 7 shows an example method 700 for predicting and validating properties of individual cells based on some embodiments of the disclosed technology.

[0078] Referring to FIG. 7, a method for predicting and validating properties of individual cells comprises: at 710, acquiring, by at least one processor, image data of a mixture of cells and markers using an imaging flow cytometer; at 720, identifying, by the at least one processor, an individual cell in the mixture on a cell placement or collection platform by dispensing the mixture onto the cell placement or collection platform after acquiring the image data, and matching a first marker sequence derived from the image data with a second marker sequence obtained from the cell placement or collection platform; at 730, inputting, by the at least one processor, the image data into a machine learning model configured to generate a predicted change in one or more properties of the identified individual cell; at 740. determining, by the at least one processor, a change in the one or more properties of the identified individual cell on the cell placement or collection platform by measuring the one or more properties of the identified individual cell at a first time point and a second time point subsequent to the first time point and comparing measurements corresponding to the first and the second time points to identify the change in the one or more properties of the identified individual cell; and. at 750, updating, by the at least one processor, the machine learning model by comparing the predicted change with the determined change to generate a comparison result and incorporating the comparison result into the machine learning model.PCT ApplicationAttorney Docket No.: 009062.8568.WOOO

[0079] FIG. 8 shows an example method 800 for determining a history of an individual cell based on some embodiments of the disclosed technology.

[0080] Referring to FIG. 8, a method for determining a history of an individual cell comprises: at 810, acquiring, by at least one processor, image data of a mixture of cells and markers using an imaging flow cytometer; at 820, identifying, by the at least one processor, an individual cell in the mixture on a cell placement or collection platform by dispensing the mixture onto the cell placement or collection platform after acquiring the image data, and matching a first marker sequence derived from the image data with a second marker sequence obtained from the cell placement or collection platform; at 830, inputting, by the at least one processor, the image data into a machine learning model configured to classify the identified individual cell according to a cellular history of the identified individual cell; and, at 840, determining, by the at least one processor, based on an output of the machine learning model, the cellular history of the identified individual cell.

[0081] FIG. 9 shows an example system for predicting properties of individual cells based on some embodiments of the disclosed technology.

[0082] Referring to FIG. 9, a system 900 for analyzing property changes of individual cells over time comprises: an imaging flow cytometer 910 configured to capture an image of a mixture of cells and markers; a cell placement or collection platform 920 configured to receive the mixture of cells and markers dispensed from the imaging flow cytometer; and at least one processor 930 operably connected to the imaging flow cytometer and the cell placement or collection platform, the at least one processor configured to: acquire image data of the mixture of cells and markers from the captured image; identify an individual cell in the mixture on the cell placement or collection platform by matching a first marker sequence derived from the image data with a second marker sequence obtained from the cell placement or collection platform; input the image data into a machine learning model configured to generate a predicted change in one or more properties of the identified individual cell; determine a change in the one or more properties of the identified individual cell on the cell placement or collection platform by measuring the one or more properties of the identified individual cell at a first time point and a second time point subsequent to the first time point and comparing measurements corresponding to the first and the second time points to identify the change in the one or more properties of the identified individual cell; andPCT ApplicationAttorney Docket No.: 009062.8568.WOOO update the machine learning model by comparing the predicted change with the determined change to generate a comparison result and incorporating the comparison result into the machine learning model.

[0083] In some embodiments, the system 900 may further include one or more memory units at least one processor operably connected to one or more memory units. The memory stores computer-readable instructions which, when executed by the at least one processor, cause the system to perform one or more of the operations described herein, including: acquiring image data from an imaging flow cytometer, determining spatial positions of individual cells on a cell placement or collection platform by comparing marker sequences, inputting image data into a trained machine learning model, predicting one or more cellular properties, and updating the machine learning model using validated data derived from temporal measurements.

[0084] In some embodiments, a cell analysis method may include acquiring an image of a mixture of cells and markers using a three-dimensional imaging flow cytometer (3D-IFC), determining a first marker sequence from the acquired image and a second marker sequence after dispensing the cells and markers onto a cell placement platform (CPP); comparing the two sequences to associate each cell image with its physical location on the platform, measuring protein expression levels of the individual cells at two different time points and calculating the relative protein production rate; generating a histogram of the protein production rates to categorize cells into high-expression and low-expression groups, and training a machine learning model, such as a convolutional autoencoder-based classifier, using image data and the derived labels, thereby enabling property prediction of unlabeled individual cells.

[0085] In some implementations, the image of the mixture of cells and markers is acquired using an imaging flow cytometer (IFC). In some implementations, the imaging flow cytometer comprises a three-dimensional imaging flow cytometer (3D-IFC) configured to generate sidescattering and transmission images of the cells.

[0086] In some implementations, dispensing the cells and markers onto the cell placement platform (CPP) comprises using a position-tracked cell placement platform. In some implementations, the position-tracked cell placement platform comprises a three-axis motorized stage and a holder configured to maintain spatial correspondence of dispensed cells.

[0087] In some implementations, the property prediction is performed using a machinePCT ApplicationAttorney Docket No.: 009062.8568.WOOO learning model trained on image data and corresponding protein production rates. In some implementations, the machine learning model comprises a convolutional autoencoder-based classifier (CAE) architecture.

[0088] Therefore, various implementations of features of the disclosed technology can be made based on the above disclosure, including the examples listed below.

[0089] Example 1. A method for predicting and validating properties of individual cells, comprising: acquiring, by at least one processor, image data of a mixture of cells and markers using an imaging flow cytometer; identifying, by the at least one processor, an individual cell in the mixture on a cell placement or collection platform by dispensing the mixture onto the cell placement or collection platform after acquiring the image data, and matching a first marker sequence derived from the image data with a second marker sequence obtained from the cell placement or collection platform; inputting, by the at least one processor, the image data into a machine learning model configured to generate a predicted change in one or more properties of the identified individual cell; determining, by the at least one processor, a change in the one or more properties of the identified individual cell on the cell placement or collection platform by measuring the one or more properties of the identified individual cell at a first time point and a second time point subsequent to the first time point and comparing measurements corresponding to the first and the second time points to identify the change in the one or more properties of the identified individual cell; and updating, by the at least one processor, the machine learning model by comparing the predicted change with the determined change to generate a comparison result and incorporating the comparison result into the machine learning model.

[0090] Example 2. The method of example 1, wherein the imaging flow cytometer includes a three-dimensional imaging flow cytometer (3D-IFC) configured to generate at least a side scattering image and a transmission image for each individual cell.

[0091] Example 3. The method of example 1, wherein the cell placement or collection platform includes a position-tracked cell placement platform (CPP).

[0092] Example 4. The method of example 1 , wherein the markers are beads.

[0093] Example s. The method of example 1, wherein the markers are fluorescently labelled cells.PCT ApplicationAttorney Docket No.: 009062.8568.WOOO

[0094] Example 6. The method of example 1 , wherein the machine learning model comprises a fused convolutional autoencoder (CAE) including a two-dimensional CAE and a three- dimensional, each configured to receive a single modality image of the individual cell as input.

[0095] Example 7. The method of example 1, wherein the one or more properties include a protein expression level of the individual cell.

[0096] Example 8. The method of example 1, wherein the one or more properties include a response to a cellular stress condition.

[0097] Example 9. The method of example 8. wherein the cellular stress condition includes at least one of thermal stress or nutrient deprivation.

[0098] Example 10. The method of example 1, wherein updating the machine learning model comprises retraining the model using the determined change as ground truth labeled data.

[0099] Example 11. The method of example 1, wherein measuring the one or more properties at each of the first and second time points is performed using fluorescence microscopy.

[0100] Example 12. The method of example 1, wherein the machine learning model is trained to classify individual cells into two or more groups based on the one or more properties of the identified individual cell.

[0101] Example 13. A system for analyzing property changes of individual cells over time, comprising: an imaging flow cytometer configured to capture an image of a mixture of cells and markers; a cell placement or collection platform configured to receive the mixture of cells and markers dispensed from the imaging flow cytometer; and a processor operably connected to the imaging flow cytometer and the cell placement or collection platform, the processor configured to: acquire image data of the mixture of cells and markers from the captured image; identify an individual cell in the mixture on the cell placement or collection platform by matching a first marker sequence derived from the image data with a second marker sequence obtained from the cell placement or collection platform; input the image data into a machine learning model configured to generate a predicted change in one or more properties of the identified individual cell; determine a change in the one or more properties of the identified individual cell on the cell placement or collection platform by measuring the one or more properties of the identified individual cell at a first time point and a second time point subsequent to the first time point andPCT ApplicationAttorney Docket No.: 009062.8568.WOOO comparing measurements corresponding to the first and the second time points to identify the change in the one or more properties of the identified individual cell; and update the machine learning model by comparing the predicted change with the determined change to generate a comparison result and incorporating the comparison result into the machine learning model.

[0102] Example 14. The system of example 13, wherein the imaging flow cytometer includes a three-dimensional imaging flow cytometer (3D-IFC) configured to generate at least a side scattering image and a transmission image for each individual cell.

[0103] Example 15. The system of example 13, wherein the cell placement or collection platform includes a position-tracked cell placement platform (CPP).

[0104] Example 16. The system of claim 13, wherein the markers are beads.

[0105] Example 17. The system of claim 13. wherein the markers are fluorescently labelled cells.

[0106] Example 18. The system of example 13. wherein the machine learning model comprises a fused convolutional autoencoder (CAE) including a two-dimensional CAE and a three-dimensional, each configured to receive a single modality image of the individual cell as input.

[0107] Example 19. The system of example 13, wherein the one or more properties include a protein expression level of the individual cell.

[0108] Example 20. The system of example 13, wherein the one or more properties include a response to a cellular stress condition.

[0109] Example 21. The system of example 20, wherein the cellular stress condition includes at least one of thermal stress or nutrient deprivation.

[0110] Example 22. The system of example 13, wherein updating the machine learning model comprises retraining the model using the determined change as ground truth labeled data.

[0111] Example 23. The system of example 13. wherein measuring the one or more properties at each of the first and second time points is performed using fluorescence microscopy.

[0112] Example 24. The system of example 13, wherein the machine learning model is trained to classify individual cells into two or more groups based on the one or more properties of thePCT ApplicationAttorney Docket No.: 009062.8568.WOOO identified individual cell.

[0113] Example 25. A method for determining a history of an individual cell, comprising: acquiring, by at least one processor, image data of a mixture of cells and markers using an imaging flow cytometer; identifying, by the at least one processor, an individual cell in the mixture on a cell placement or collection platform by dispensing the mixture onto the cell placement or collection platform after acquiring the image data, and matching a first marker sequence derived from the image data with a second marker sequence obtained from the cell placement or collection platform; inputting, by the at least one processor, the image data into a machine learning model configured to classify the identified individual cell according to a cellular history of the identified individual cell; and determining, by the at least one processor, based on an output of the machine learning model, the cellular history of the identified individual cell.

[0114] Example 26. The method of example 25, wherein further comprising: training, by the at least one processor, the machine learning model using training data comprising: image data of cells acquired at a first time point before any change is observable; and ground truth labels generated by applying one or more conditions affecting the cells and determining whether one or more properties of the cells change due to the one or more conditions based on measurements obtained at a second time point subsequent to the first time point.

[0115] Example 27. The method of example 25, wherein the imaging flow cytometer includes a three-dimensional imaging flow cytometer (3D-IFC) configured to generate at least a side scattering image and a transmission image for each individual cell.

[0116] Example 28. The method of example 25, wherein the cell placement or collection platform includes a position-tracked cell placement platform (CPP).

[0117] Example 29. The method of example 25, wherein the markers are beads.

[0118] Example 30. The method of example 25, wherein the markers are fluorescently labelled cells.

[0119] Example 31. The method of example 25, wherein the machine learning model comprises a fused convolutional autoencoder (CAE) including a two-dimensional CAE and a three-dimensional, each configured to receive a single modality image of the individual cell as input.PCT ApplicationAttorney Docket No.: 009062.8568.WOOO

[0120] Example 32. The method of example 25, wherein the cellular history of the identified individual cell includes whether the identified individual cell has stressed, intoxicated, or nutrient deprived.

[0121] It is understood that the various disclosed embodiments may be implemented individually, or collectively, using devices comprised of various optical components, electronics hardware and / or software modules and components. These devices, for example, may comprise a processor, a memory unit, an interface that are communicatively connected to each other, and may range from desktop and / or laptop computers, to mobile devices and the like. The processor and / or controller can perform various disclosed operations based on execution of program code that is stored on a storage medium. The processor and / or controller can, for example, be in communication with at least one memory and with at least one communication unit that enables the exchange of data and information, directly or indirectly, through the communication link with other entities, devices and networks. The communication unit may provide wired and / or wireless communication capabilities in accordance with one or more communication protocols, and therefore it may comprise the proper transmitter / receiver antennas, circuitry and ports, as well as the encoding / decoding capabilities that may be necessary for proper transmission and / or reception of data and other information.

[0122] Various information and data processing operations described herein may be implemented in one embodiment by a computer program product, embodied in a computer- readable medium, including computer-executable instructions, such as program code, executed by computers in networked environments. A computer-readable medium may include removable and non-removable storage devices including, but not limited to, Read Only Memory (ROM), Random Access Memory (RAM), compact discs (CDs), digital versatile discs (DVD), etc. Therefore, the computer-readable media that is described in the present application comprises non-transitory storage media. Generally, program modules may include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Computer-executable instructions, associated data structures, and program modules represent examples of program code for executing steps of the methods disclosed herein. The particular sequence of such executable instructions or associated data structures represents examples of corresponding acts for implementing the functions described in such steps or processes.PCT ApplicationAttorney Docket No.: 009062.8568.WOOO

[0123] The foregoing description of embodiments has been presented for purposes of illustration and description. The foregoing description is not intended to be exhaustive or to limit embodiments of the present invention to the precise form disclosed, and modifications and variations are possible in light of the above teachings or may be acquired from practice of various embodiments. The embodiments discussed herein were chosen and described in order to explain the principles and the nature of various embodiments and its practical application to enable one skilled in the art to utilize the present invention in various embodiments and with various modifications as are suited to the particular use contemplated. While operations are depicted in the drawings in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. The features of the embodiments described herein may be combined in all possible combinations of methods, apparatus, modules, and systems.

[0124] Only a few implementations and examples are described and other implementations, enhancements and variations can be made based on what is described and illustrated in this patent document.

Claims

PCT ApplicationAttorney Docket No.: 009062.8568.WOOOCLAIMSWhat is claimed is:

1. A method for predicting and validating properties of individual cells, comprising: acquiring, by at least one processor, image data of a mixture of cells and markers using an imaging flow cytometer; identifying, by the at least one processor, an individual cell in the mixture on a cell placement or collection platform by dispensing the mixture onto the cell placement or collection platform after acquiring the image data, and matching a first marker sequence derived from the image data with a second marker sequence obtained from the cell placement or collection platform; inputting, by the at least one processor, the image data into a machine learning model configured to generate a predicted change in one or more properties of the identified individual cell; determining, by the at least one processor, a change in the one or more properties of the identified individual cell on the cell placement or collection platform by measuring the one or more properties of the identified individual cell at a first time point and a second time point subsequent to the first time point and comparing measurements corresponding to the first and the second time points to identify the change in the one or more properties of the identified individual cell; and updating, by the at least one processor, the machine learning model by comparing the predicted change with the determined change to generate a comparison result and incorporating the comparison result into the machine learning model.

2. The method of claim 1, wherein the imaging flow cytometer includes a three- dimensional imaging flow cytometer (3D-IFC) configured to generate at least a side scattering image and a transmission image for each individual cell.

3. The method of claim 1, wherein the cell placement or collection platform includes a position-tracked cell placement platform (CPP).PCT ApplicationAttorney Docket No.: 009062.8568.WOOO4. The method of claim 1 , wherein the markers are beads.

5. The method of claim 1, wherein the markers are fluorescently labelled cells.

6. The method of claim 1, wherein the machine learning model comprises a fused convolutional autoencoder (CAE) including a two-dimensional CAE and a three-dimensional, each configured to receive a single modality image of the individual cell as input.

7. The method of claim 1, wherein the one or more properties include a protein expression level of the individual cell.

8. The method of claim 1, wherein the one or more properties include a response to a cellular stress condition.

9. The method of claim 8, wherein the cellular stress condition includes at least one of thermal stress or nutrient deprivation.

10. The method of claim 1, wherein updating the machine learning model comprises retraining the model using the determined change as ground truth labeled data.

11. The method of claim 1, wherein measuring the one or more properties at each of the first and second time points is performed using fluorescence microscopy.

12. The method of claim 1, wherein the machine learning model is trained to classify individual cells into two or more groups based on the one or more properties of the identified individual cell.

13. A system for analyzing property changes of individual cells over time, comprising: an imaging flow cytometer configured to capture an image of a mixture of cells and markers;PCT Application Attorney Docket No.: 009062.8568.WOOO a cell placement or collection platform configured to receive the mixture of cells and markers dispensed from the imaging flow cytometer; and at least one processor operably connected to the imaging flow cytometer and the cell placement or collection platform, the at least one processor configured to: acquire image data of the mixture of cells and markers from the captured image; identify an individual cell in the mixture on the cell placement or collection platform by matching a first marker sequence derived from the image data with a second marker sequence obtained from the cell placement or collection platform; input the image data into a machine learning model configured to generate a predicted change in one or more properties of the identified individual cell; determine a change in the one or more properties of the identified individual cell on the cell placement or collection platform by measuring the one or more properties of the identified individual cell at a first time point and a second time point subsequent to the first time point and comparing measurements corresponding to the first and the second time points to identify the change in the one or more properties of the identified individual cell; and update the machine learning model by comparing the predicted change with the determined change to generate a comparison result and incorporating the comparison result into the machine learning model.

14. The system of claim 13, wherein the imaging flow cytometer includes a three- dimensional imaging flow cytometer (3D-IFC) configured to generate at least a side scattering image and a transmission image for each individual cell.

15. The system of claim 13, wherein the cell placement or collection platform includes a position-tracked cell placement platform (CPP).

16. The system of claim 13, wherein the markers are beads.

17. The system of claim 13, wherein the markers are fluorescently labelled cells.PCT ApplicationAttorney Docket No.: 009062.8568.WOOO18. The system of claim 13, wherein the machine learning model comprises a fused convolutional autoencoder (CAE) including a two-dimensional CAE and a three-dimensional, each configured to receive a single modality image of the individual cell as input.

19. The system of claim 13, wherein the one or more properties include a protein expression level of the individual cell.

20. The system of claim 13, wherein the one or more properties include a response to a cellular stress condition.

21. The system of claim 20, wherein the cellular stress condition includes at least one of thermal stress or nutrient deprivation.

22. The system of claim 13, wherein updating the machine learning model comprises retraining the model using the determined change as ground truth labeled data.

23. The system of claim 13, wherein measuring the one or more properties at each of the first and second time points is performed using fluorescence microscopy.

24. The system of claim 13, wherein the machine learning model is trained to classify individual cells into two or more groups based on the one or more properties of the identified individual cell.

25. A method for determining a history of an individual cell, comprising: acquiring, by at least one processor, image data of a mixture of cells and markers using an imaging flow cytometer; identifying, by the at least one processor, an individual cell in the mixture on a cell placement or collection platform by dispensing the mixture onto the cell placement or collection platform after acquiring the image data, and matching a first marker sequence derived from the image data with a second marker sequence obtained from the cell placement or collection platform;PCT ApplicationAttorney Docket No.: 009062.8568.WOOO inputting, by the at least one processor, the image data into a machine learning model configured to classify the identified individual cell according to a cellular history of the identified individual cell; and determining, by the at least one processor, based on an output of the machine learning model, the cellular history of the identified individual cell.

26. The method of claim 25, wherein further comprising: training, by the at least one processor, the machine learning model using training data comprising: image data of cells acquired at a first time point before any change is observable; and ground truth labels generated by applying one or more conditions affecting the cells and determining whether one or more properties of the cells change due to the one or more conditions based on measurements obtained at a second time point subsequent to the first time point.

27. The method of claim 25, wherein the imaging flow cytometer includes a three- dimensional imaging flow cytometer (3D-IFC) configured to generate at least a side scattering image and a transmission image for each individual cell.

28. The method of claim 25, wherein the cell placement or collection platform includes a position-tracked cell placement platform (CPP).

29. The method of claim 25, wherein the markers are beads.

30. The method of claim 25, wherein the markers are fluorescently labelled cells.

31. The method of claim 25, wherein the machine learning model comprises a fused convolutional autoencoder (CAE) including a two-dimensional CAE and a three-dimensional, each configured to receive a single modality image of the individual cell as input.PCT Application Attorney Docket No.: 009062.8568.WOOO32. The method of claim 25, wherein the cellular history of the identified individual cell includes whether the identified individual cell has stressed, intoxicated, or nutrient deprived.

Images