WO2025199290A1 - Methods for rapid assessing t cells health and predicting optimal growth condition thereof - Google Patents

Abstract

The present disclosure provides, inter alia, methods for rapid assessing the health of a cell including T cells and methods for predicting the optimal growth condition for cell, such as T cells, expansion. Also provided are methods for treating or ameliorating the effects of a disease in a subject in need thereof, with cells assessed and/or expanded using the methods disclosed herein.

Description

METHODS FOR RAPID ASSESSING T CELLS HEALTH AND PREDICTING OPTIMAL GROWTH CONDITION THEREOF

CROSS REFERENCE TO RELATED APPLICATIONS

[0001] The present application claims benefit of U.S. Provisional Patent Application Serial No. 63/567,743, filed on March 20, 2024, U.S. Provisional Patent Application Serial No. 63/632,323, filed on April 10, 2024, U.S. Provisional Patent Application Serial No. 63/638,189, filed on April 24, 2024, and U.S. Provisional Patent Application Serial No. 63/703,168, filed on October 3, 2024. The entire contents of the aforementioned applications are hereby incorporated by reference.

FIELD OF DISCLOSURE

[0002] The present disclosure provides, inter alia, methods for rapid assessing the health of T cells and methods for predicting the optimal growth condition for T cells expansion.

GOVERNMENT FUNDING

[0003] This invention was made with government support under AI154867, CA013696, and AI110593 awarded by the National Institutes of Health, and 1743420 awarded by the National Science Foundation. The government has certain rights in the invention.

BACKGROUND

[0004] T cells have shown particular promise as a “living drug”, as exemplified by the emergence of chimeric antigen receptor (CAR) T cell therapy to treat cancer. This therapy involves genetic modification of patients’ T cells. Specifically, T cells are isolated from patients’ blood through a process called leukapheresis, and the collected T cells are engineered in lab to express a chimeric antigen receptor (CAR) on their surface recognizing a specific antigen present on the surface of cancer cells. This engineered receptor contains an extracellular domain for recognition, a transmembrane domain, and an intracellular signaling domain that activates T cells upon antigen binding. The genetically modified T cells are then expanded in lab to reach a clinically relevant number and are reinfused to patients. Once in the body, these cells circulate to cancer cells expressing the target antigen and initiate a robust immune response against the cancer.

[0005] CAR T cell therapy represents a paradigm shift in cancer treatment and shows many advantages including targeted action, long-lasting effects, and personalized treatment. CAR T cell therapy has been particularly effective in treating certain types of cancers, particularly hematologic malignancies like leukemia and lymphoma. CAR T cell therapy has shown remarkable success in treating pediatric and adult patients with relapsed or refractory acute lymphoblastic leukemia (ALL).

[0006] While CAR T cells have resulted in a dramatic treatment response for patients with hematologic malignancies, many challenges remain in the production of functional and stable T cells necessary for adoptive therapy. For example, T cells from patients being treated for chronic lymphocytic leukemia (CLL) exhibit impaired expansion, which complicates CAR-T therapy applied in these patients.

[0007] Accordingly, there is a need for developing methods for assessing the health of T cells and optimizing T cells expansion. This disclosure is directed to meeting these and other needs.

SUMMARY

[0008] The present disclosure relates to a comprehensive approach integrating biomaterials, imaging, and deep learning to assess the health of T cells and predict cell culture conditions that optimize expansion. This disclosure developed a framework encompassing a short-term T cell spreading assay coupled with a deep learning model that outperforms morphology-based analysis. The objective is to predict the optimal growth condition T cell expansion through a rapid test, which can more efficiently make use of patient’s T cells, thereby increasing the successful rate of CAR-T therapy. As a second objective, this model provides a measure of T cell functionality that will be used to guide cellular immunotherapy.

[0009] The framework disclosed herein demonstrated the efficacy in classifying T cell images into Healthy or chronic lymphocytic leukemia (CLL) categories with an impressive AUC of 99.6%. Moreover, the framework was also able to classify the stiffness of substrates T cells are encountering into Soft or Hard with an AUC of 99.5%. Building upon these results, the next phase involves training a regression model to predict the proliferation index, a parameter typically acquired through a 2-week trial, using short-term T-cell spreading images as input. By providing proliferation index predictions within hours, our approach offers a significant advancement in expediting the optimization of cell culture conditions.

[0010] Accordingly, one embodiment of the present disclosure is a method for rapid assessing the health of a cell from a subject in need thereof. This method comprises the steps of: (a) training a model that provides a standard reference indicating the health of the cell, comprising: (i) isolating the cell from a healthy subject and expanding them on a functionalized substrate; (ii) obtaining raw images for the cell by immunofluorescence and fixed imaging; (iii) processing the raw images to generate sliding patches; (iv) measuring a number of parameters based on the sliding patches; (v) assigning an indicator of health to the cell and recording the measured values for the parameters; (vi) repeating steps (a-i) to (a-v) with the cell from a number of healthy subjects; (vii) consolidating, analyzing and adjusting the measured values for the parameters from all tested healthy subjects to obtain a classifier for the healthy cell; (viii) repeating steps (a-i) to (a-vii) by isolating the cell from a subject with a condition and expanding them on the same functional substrate, wherein the cell from the subject with the condition receives an indicator of the condition and obtains a classifier for the cell having the condition; (ix) repeating step (a-viii) by isolating the cell from a subject with a different condition and expanding them on the same functional substrate; and (x) obtaining the standard reference by consolidating and analyzing the classifiers for the healthy cell and the cell having any condition, (b) assessing the health of the cell from a subject in need thereof, comprising: (i) isolating the cell from the subject in need thereof and obtaining the measured values for the parameters as described in steps (a-i) to (a- iv); (ii) entering the measured values in step (b-i) into the trained model; and (iii) receiving the indicator of health, and (c) selecting a suitable cell expansion plan for the subject.

[0011] Another embodiment of the present disclosure is a method for predicting the optimal growth condition for expansion of a cell from a subject in need thereof. This method comprises the steps of: (a) training a model that provides a standard reference predicting the optimal growth condition for expansion of the cell, comprising: (i) isolating the cell from a healthy subject and expanding them on a functionalized substrate; (ii) obtaining raw images for the cell by immunofluorescence and fixed imaging; (iii) processing the raw images to generate sliding patches; (iv) measuring a number of parameters based on the sliding patches; (v) assigning an expansion indicator to the cell and recording the measured values for the parameters; (vi) repeating steps (a-i) to (a-v) with the cell from a number of healthy subjects; (vii) consolidating, analyzing and adjusting the measured values for the parameters from all tested healthy subjects to obtain an expansion classifier for the healthy cell on the functionalized substrate; (viii) repeating steps (a-i) to (a-vii) by expanding the cell from healthy subjects on a different functionalized substrate to receive a different expansion indicator and obtain an expansion classifier on the different functionalized substrate; (ix) changing the functionalized substrate and repeating steps (a-viii); (x) repeating steps (a-i) to (a-ix) by isolating the cell from a subject with a condition; (xi) repeating step (a-x) by isolating the cell from a subject with a different condition; and (xii) obtaining the standard reference by consolidating and analyzing the classifiers for the healthy cell and the cell having any condition, (b) predicting the optimal growth condition for expansion of the cell from a subject in need thereof, comprising: (i) isolating the cell from the subject in need thereof and obtaining the measured values for the parameters as described in steps (a-i) to (a-iv); (ii) entering the measured values in step (b-i) into the trained model; and (iii) receiving the expansion indicator, and (c) selecting an optimal cell expansion plan for the subject.

[0012] A further embodiment of the present disclosure is a method for treating or ameliorating the effects of a disease in a subject in need thereof. This method comprises: (a) assessing the health of T cells isolated from the subject according to the method disclosed herein; (b) selecting an optimal growth condition for T cells expansion for the subject according to the method disclosed herein; (c) expanding T cells for the subject under the optimal growth condition of step (b) to a clinically relevant number; (d) reinfusing the expanded T cells to the subject; (e) after a course of treatment, reassessing the health of T cells in the subject by repeating step (a); (f) if the reassessment result in step (e) is acceptable, continuing expanding T cells under the current optimal growth condition, or if the reassessment result in step (e) is unacceptable, adjusting T cells expansion by selecting a different optimal growth condition by repeating step (b) and expanding T cells under the adjusted optimal growth condition; (g) reinfusing the expanded T cells in step (f) to the subject; and (h) repeating steps (e) to (g) as necessary.

BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

[0014] Figures 1A-1E show clustering analysis of cells from CLL patients revealed three Groups that describe proliferative potential. (A) Scree plot indicating the percentage of explained variance associated with each Dimension of a sevenfactor FAMD analysis. Subsequent analysis focused on Dim1 + Dim2, which explains over 50% of variance. (B) Contribution of each factor to Dim1 + Dim2. The red line indicates 14.3%, a threshold representing equal contribution by each factor. (C) Analysis by k-medoids clustering using factors with contributions above the threshold indicated in (B) produced three Groups, which are coded in this FAMD plot showing Dim1 and Dim2. (D) Maximum doublings varied as a function of Group assignment. (E) PD-1 , Alignment, and IL-2 secretion as a function of Group assignment. In all panels, data are mean ± SD *P < 0.05, **P < 0.005, ***P < 0.0005, ****P < 0.0001 , using ANOVA and Tukey tests. All comparisons that were significant at a = 0.05 are indicated in this figure. Open symbols represent conditions for which missing data was imputed.

[0015] Figures 2A-2C show the glass-supported Poly (dimethyl siloxane) (PDMS) substrate preparation. (A) Blending commercially available Sylgard 527 and Sylgard 184 at different ratios can fabricate PDMS with varying stiffnesses. (B) Thin PDMS layers were created on glass coverslips (thickness #0, Electron Microscopy Sciences, Hatfield, PA, USA). A droplet of PDMS mixture was pressed using a PDMS cube to flatten the droplet and create a thin (~20 pm) layer, which would be peeled off after curing overnight at 65 °C. The PDMS cubes were silanized overnight with (tridecafluoro-1 , 1 ,2, 2, -tetrahydrooctyl)-1 -trichlorosilane (United Chemical Technologies, Bristol, PA, USA), to facilitate removal from the PDMS substrates. (C) Mechanonsensing evaluation under different anti-CD3: anti-CD28 ratios. Ratio of 1 :1 shows the most significant mechanosesning reflected by cell area. [0016] Figures 3A-3C show the activating antibody coating and validation. (A) Visualization of anti-CD3 and anti-CD28 on PDMS surface, with half the antibodies labeled with Alexa Fluor 568 NHS Ester (Succinimidyl Ester). (B) Quantification of antibody across stiffness to make sure the antibody presentation is the same. (C) Validation of antibody-coated PDMS. T cells can only attach and spread on top of antibody-coated PDMS.

[0017] Figure 4 shows the workflow of cell spreading assay. T cells are isolated from blood and are allowed to spread on glass-supported PDMS thin film for 40 min, followed by fixation, permeabilization, immunostaining, and imaging.

[0018] Figures 5A-5B show the dynamics of T cell spreading on PDMS substrates. (A) Live imaging captures T cell spreading over 60 min and identifies 40 min as a time point during which cells stabilize on each surface and produce the greatest difference as a function of elastic modulus. (B) Fixed imaging analysis comparing 20 min, 30 min, and 40 min timepoints. The 40 min timepoint retains the stability and resolution seen in the live-cell assays. Each data point stands for each cell. Statistical significance was determined using one-way ANOVA with Tukey multiple comparison test, **p<0.01 , ***p<0.05, ****p<0.001.

[0019] Figures 6A-6C show the fixed imaging under 100x and 40x magnification. (A) Healthy T cells were imaged under 100x magnification with actin stained. (B) Healthy and CLL T cells were imaged under 40x magnification with actin stained. More cells can be visualized and analyzed under 40x magnification with imaging quality guaranteed. (C) Measurement of cell area under 40x and 100x magnification shows no significant difference.

[0020] Figures 7A-7D show the characterization of PDMS substrates and visualization of T cell spreading from both healthy donors and CLL patients. (A) Schematic of antibody-coated PDMS thin layer to activate T cells. (B) Indentation testing was performed to measure the Young’s modulus of different PDMS formulations, with varying mass ratios of Sylgard 527 and Sylgard 184. Data are mean ± s.d., n = 4 for 10:1 (250 kPa), n = 3 for the other formulations. (C) Quantification of antibody coating indicates a consistent level of OKT3 and 9.3 coated on the surfaces across different formulations of PDMS. Data are mean ± s.d., n = 4 samples for each stiffness condition, ns: p>0.05. (D) Fixed imaging finds that CLL T cells exhibit a smaller spreading area and a higher roundness than Healthy T cells, supporting the concept that disease state affects T cell morphology. Scale bar: 20 pm.

[0021] Figures 8A-8B show the quantitative analysis of T cell Area and Roundness from Healthy donors and CLL patients across three stiffness conditions. (A) CLL T cells show significantly smaller Area and higher Roundness than Healthy donors, and this applies to all three stiffness conditions. Data are mean ± s.d., each data point represents an individual substrate consisting of approximately 100 cells. Different symbols reflect different conditions: Healthy or CLL. Statistical significance was determined using unpaired t test with Welch’s correction across all cells captured for each condition, **** p < 0.001. (B) T cells from healthy donors and CLL patients respond to substrate stiffness. Data are mean ± s.d., each data point represents an individual substrate consisting of approximately 100 cells. Statistical significance was determined using two-way ANOVA followed by Tukey multiple comparison test across all cells captured for each condition, * p < 0.05, ** p < 0.01 , *** p < 0.005, **** p < 0.001.

[0022] Figures 9A-9B show the donor-to-donor variation exists in T cell mechanosensing. (A) T cell spreading area as a function of stiffness for 3 healthy donors. (B) T cell spreading area as a function of stiffness for 6 CLL patients.

[0023] Figures 10A-10B show the quantitative analysis of T cell Area over 60 min live imaging from Healthy donors and CLL patients across three stiffness conditions. (A) Time-lapse T cell spreading from Healthy donor H1. (B) Time-lapse T cell spreading from CLL patients D18 and D51.

[0024] Figure 11 shows the imaging processing workflow. Fixed imaging was performed under 40X magnification to acquire more cells in a field of view for analysis. Image analysis was performed in ImageJ using functions of Smoothing, Thresholding, Set Measurement, and Analyze Particles to measure morphological features of single cells, including Area, Perimeter, Width, Height, Major, Minor, Circularity, Feret’s Diameter, Aspect Ratio (AR), Roundness, and Solidity. Only accurately segmented single cells were selected by quality control for further analysis. [0025] Figures 12A-12B show that PCA reveals the variance between CLL and Healthy T cells and identifies important morphological features contributing to the variance. (A) Two-dimensional representation of PCA analysis. Projection of the data along PC1 showed a separation between Healthy (blue) and CLL (red). Three stiffness conditions which the data were derived from were also shape-coded. Each data point represents an individual sample. (B) Feature importance on PC1 and PC2.

[0026] Figures 13A-13B show the effect of cytoskeletal protein inhibitors on T cell mechanosensing. (A) T cells from a healthy donor were treated with DMSO control, CK666 (100 pM), or Y-27632 (60 pM) for 15 min before being seeded onto PDMS substrates, followed by fixation, permeabilization, and actin staining. Image examples (250 kPa substrate) were shown; scale bar: 10 pm. (B) Quantitative analysis reveals the effect of CK666 and Y-27632. Data are mean ± s.d. For DMSO, n = 10; for CK666, n = 8; for Y-27632, n - 4. Different symbols reflect different stiffness conditions. Statistical significance was determined using two-way ANOVA with Tukey multiple comparison test, * p < 0.05, **** p < 0.001 .

[0027] Figures 14A-14B show the effect of ICAM-1 on T cell mechanosensing. (A) T cell spreading area across stiffness with and without ICAM- 1. ICAM-1 did not affect T cell mechanosensing. (B) Increased spreading area in presence of ICAM. This experiment was conducted with one healthy donor’s samples. Data are mean ± s.d. N=100 cells for each bar. Statistical significance was determined using one-way ANOVA with Tukey multiple comparison test, * p < 0.05, ** p < 0.01 , *** p < 0.005, **** p < 0.001 .

[0028] Figures 15A-15C show that T cell IL-2 secretion shows sensitivity to subsrate stiffness. (A) T cell IL-2 secretion at 3hr and 4hr. By 4hr, T cells have had sufficient time to complete the secretion of IL-2 and mechanosensing manifest at this time point. (B) Nocodazole (NZ) significantly decreased the percentage of cells secreting IL-2, and Taxol (TX) slightly decreased the percentage. (C) Both NZ and TX significantly impaired the overall IL-2 secretion level of cells, and mechanonsensing disappeared under both treatments.

[0029] Figures 16A-16B show the demonstration of indentation result by Instron. (A) Load over time plot reflecting 6 blocks, with 3 cycles of loading 10s and holding 20s. (B) Displacement overweight plot to find the slope after linear regression to calculate young’s modulus.

[0030] Figure 17 shows the workflow of image-based deep learning to classify Healthy and CLL. In preprocessing, each raw image (1002x1004) was broken into 25 (5 rows x 5 columns) smaller image patches (224x224). The image patches were then split into Train, Validation, and Test datasets at a ratio of 80%:10%:10%. Image patches were input into a pretrained Swin Transformer model for classification into Healthy and CLL and original image-level prediction was decided after majority voting mechanism.

[0031] Figures 18A-18C show the pretrained Swin Transformer model performance for Healthy vs. CLL classification on hard PDMS substrate, with unfrozen weights. (A) Learning curve showing the train/validation loss and accuracy across 100 epochs. (B) ROC was generated for the test dataset, with an AUC of 0.887. (C) Confusion matrix generated for test dataset.

[0032] Figures 19A-19C show the un-pretrained Swin Transformer model performance for Healthy vs. CLL classification on hard surface. (A) Learning curve showing the train/validation loss and accuracy across 100 epochs. (B) ROC was generated for the test dataset, with an AUC of 0.783. (C) Confusion matrix generated for test dataset.

[0033] Figures 20A-20C show the pretrained Swin Transformer model performance for Healthy vs. CLL classification on hard surface, freezing the feature extraction weights. (A) Learning curve showing the train/validation loss and accuracy across 100 epochs. (B) ROC was generated for the test dataset, with an AUC of 0.887. (C) Confusion matrix generated for test dataset.

[0034] Figures 21A-21C show the Pretrained ResNet-50 model performance for Healthy vs. CLL classification on hard surface, unfreezing the feature extraction weights. (A) Learning curve showing the train/validation loss and accuracy across 100 epochs. (B) ROC was generated for the test dataset, with an AUC of 0.845. (C) Confusion matrix generated for test dataset.

[0035] Figures 22A-22B show the comparison of Examples 1 and 2 in the results of cell area as a function of substrate stiffness. (A) Example 1 used a mixture of Sylgard 527 and Sylgard 184 and a direct coating method. There were 3 healthy donors and 6 CLL patients included. (B) Example 2 used only Sylgard 184 and tuned the stiffness by changing the crosslinker to base ratio. New coating method was used and new donors (3 Healthy donors and 7 CLL patients) were included.

[0036] Figures 23A-23C show the characterization of PDMS substrates and antibody coating. (A) Schematic of antibody-coated PDMS thin layer to activate T cells. (B) Indentation testing was performed to measure the Young’s modulus of different PDMS formulations, with varying mass ratios of crosslinker to base of Sylgard 184. Data are mean ± s.d., n = 4 for each formulation. (C) Quantification of antibody coating indicates a consistent level of OKT3 and 9.3 coated on the surfaces across different formulations of PDMS. Data are mean ± s.d., n = 4 samples for each stiffness condition, ns: p>0.05.

[0037] Figure 24 shows the spreading area and expansion profiles of T cells from 3 healthy donors and 7 patients across three different stiffness conditions. Single feature such as Area is not sufficient to predict proliferation, reflected by the different trend between Area and Max Doubling as a function of stiffness.

[0038] Figure 25 shows the regression model workflow to predict max doublings. In the preprocessing, each raw image (1002x1004) was sliced into 25 smaller image patches (224x224), which was input to a pretrained Swin Transformer model for feature extraction. The model head was modified to output a numerical value. The loss between the output and true label was calculated by MSE() loss function. After predicting max doublings under different stiffness conditions, the stiffnesses were ranked to output the best stiffness for expansion.

[0039] Figure 26 shows the learning curves of loss and accuracy over 250 epochs for both train and validation, with the best performance observed at epoch 118.

[0040] Figure 27 shows the scatter plot showing the correlation between model predictions and true labels. The predicted value with the peak density under each condition was highlighted.

[0041] Figure 28 shows the visualization of predictions across three stiffness and comparison between predictions and ground truth in stiffness ranking based on proliferative capacity. [0042] Figure 29 shows the model performance in outputting the optimal stiffness for expansion.

[0043] Figure 30 is a schematic of workflow for studying T cell subset mechanosensing. T cells were seeded on PDMS with varied stiffness and mechanosensing in short-term spreading, and long-term expansion were evaluated. For short-term spreading, cells were allowed to interact with substrates for 40 min, then fixed, permeabilized, and stained with biomarkers to distinguish T cell subsets from imaging.

[0044] Figures 31A-31 B show the flow cytometry panel and example of gating and analysis. (A) Gating of live cells. (B) Gating of CD4+/CD8+ from CD4 staining, and gating of naive, central memory, effector memory, and effector T cells by CD45RA and CCR7 staining.

[0045] Figures 32A-32D show the flow cytometry results showing T cell composition of 3 healthy donors and 7 CLL patients. (A) Percentages of naive T cells, central memory (CM) T cells, effector memory (EM) T cells, and effector T cells out of total live T cells. (B) Percentages of CD4+ and CD8+ T cells out of total live T cells. (C) Percentage of naive T cells, central memory (CM) T cells, effector memory (EM) T cells, and effector T cells out of live CD4+ T cells. (D) Percentage of naive T cells, central memory (CM) T cells, effector memory (EM) T cells, and effector T cells out of live CD8+ T cells.

[0046] Figures 33A-33B show the T cell images from different fluorescence channels and subset characterization. (A) Actin staining for visualization of T cell morphology. T cell surface markers CD4, CD45RA, and CCR7 for characterization of single cell subtype. (B) Marker guide for T cell phenotyping.

[0047] Figure 34 shows the cell spreading area of CD4+ and CD8+ T cells on different stiffnesses. Donor to donor variation exists in terms of T cell subtype mechanosensing responses.

[0048] Figure 35 shows that the mechanosensitivity of various T cell subtypes — Naive, Central Memory (CM), Effector Memory (EM), and Effector T cells — was evaluated by analyzing their spreading areas on substrates of different stiffnesses in samples from two healthy donors and one CLL patient. DETAILED DESCRIPTION OF THE DISCLOSURE

[0049] CAR T cell therapy has shown remarkable outcomes in blood cancer treatment. However, using cells as a “living drug” still presents significant challenges, particularly in chronic cancers such as chronic lymphocytic leukemia (CLL). Despite the early success of CAR T clinical trials in CLL patients including Bill Ludwig and Doug Olson in 2010, CAR T therapy for CLL did not get FDA approval until March 2024, Bristol Myers Squibb’s Breyanzi ® as the First and Only CAR T Cell Therapy for Adults with Relapsed or Refractory Chronic Lymphocytic Leukemia (CLL) or Small Lymphocytic Lymphoma (SLL). The variability in therapy’s effectiveness stems from factors like patient-dependent immunodeficiency, which renders some patients ineligible for the treatment, or T cell exhaustion, which leads to the decreased proliferation and cytotoxicity. These limitations complicate the preparation of therapeutic quantities of cells and ensuring efficacy once infused to patients. Therefore, the ability to rapidly assess the responsiveness of an individual’s cell would greatly improve T cell production by tailoring expansion conditions, avoiding time- and resource-consuming trial-and-error approaches. The rapid test would also provide insights into T cell health throughout the course of treatment.

[0050] To address this unmet need, currently available assays include cell count, biomarkers, and cytokine secretion. However, previous work demonstrated that single biomarkers such as PD1 expression cannot reflect proliferation, while high-level cellular functions, specifically cell migration, can provide enhanced insights into long-term cell function compared to molecular measurements alone. The existing assay measures cell alignment onto activating antibodies micropatterned on glass coverslips, and combining cell alignment with other biomarkers such as PD1 expression, IL-2 secretion, Rai stage can categorize CLL patients into three groups which correspond to their proliferative capacity. However, this assay remains complex, requiring advanced microfabrication techniques and the need to test multiple markers. Additionally, as an unsupervised learning approach, it can only categorize patients, rather than predict the functional outcomes directly. Therefore, there is a need for a simple and more automated predictive model, leveraging cell-level behaviors for functional predictions. [0051] Machine learning (ML) has emerged as a powerful tool in the field of cell imaging, revolutionizing the way to analyze and interpret cellular data. The integration of machine learning with advanced imaging technologies enables automated, high-throughput, and quantitative analysis of cell images.

[0052] Cell imaging allows for the visualization and study of the complex structures and functions of cells. Imaging techniques such as fluorescence microscopy, confocal microscopy, and live-cell imaging provide detailed information about cell morphology, movement, and interactions. However, traditional analysis methods present significant challenges due to the sheer volume and complexity of imaging data.

[0053] Machine learning, particularly deep learning, offers robust solutions for automating the analysis of cell images. By leveraging large datasets and sophisticated algorithms, machine learning can extract important features from images with high accuracy and efficiency. The applications include image segmentation, feature extraction and classification, and phenotypic screening. Accurate segmentation of cells is crucial for quantitative analysis, and machine learning algorithms, such as convolutional neural networks (CNN) can automatically segment cell images, significantly reducing the time and effort compared to manual annotation. Machine learning can also extract features from cell images, such as cell shape, size, texture, and intensity, which can be used to categorize cells into different types. For example, ML algorithms can differentiate between healthy and diseased cells based on morphological differences. To improve performance, it is needed to collect larger datasets, learn more powerful models, and use better techniques for preventing overfitting.

[0054] While machine learning has significantly advanced cell image analysis, there still remain challenges. High-quality and annotated datasets are essential for training machine learning models. However, generating such datasets can be timeconsuming and labor-intensive. Besides, models trained on specific datasets may not generalize well to other datasets collected by different imaging conditions or cell types. Furthermore, machine learning models, particularly deep learning, often operate as “black boxes”, making it difficult to interpret their decision-making processes. Therefore, developing interpretable models is crucial for gaining biological insights. Despite these challenges, machine learning in cell imaging presents a bright future. Advances in computational power, algorithm development, and data availability will continue to enhance the capabilities and applications of machine learning in this field.

[0055] Decision tree is a type of supervised learning algorithm that recursively split the data into subsets based on the value of input features, creating a tree-like structure of decisions. Each internal node of the tree represents a decision based on the value of a feature, each branch represents the outcome of the decision, and each leaf node represents a final prediction or outcome. It can be used for both classification and regression tasks. Decision tree can handle both numerical and categorical data and requires little data preprocessing. It is also intuitive to understand, as the decision-making process mimics human reasoning. However, decision trees are prone to overfitting, especially when they are allowed to grow deep. Limiting tree depth can help mitigate overfitting.

[0056] Random forests are an ensemble learning method that extends the idea of decision trees to improve their performance and robustness. A random forest consists of a collection (or "forest") of decision trees, typically trained with the "bagging" method, which involves training each tree on a random subset of the data to reduce variance and prevent overfitting. In addition to bagging, random forests introduce another layer of randomness by selecting a random subset of features at each split in the decision trees. The final prediction of a random forest is based on a “crowd of wisdom” obtained by aggregating the predictions of all individual trees. For classification tasks, this is usually done through a majority vote, while for regression tasks, the average prediction of all trees is taken.

[0057] Random forests offer several advantages over single decision trees. They are more accurate and robust to overfitting, especially with high-dimensional data. They also provide a measure of feature importance, which can be useful for understanding the underlying data and for feature selection. However, the increased complexity of random forests comes at the cost of reduced interpretability compared to a single decision tree. Additionally, random forests can be computationally intensive, especially with a large number of trees and high-dimensional data. [0058] Computer vision modeling has long been dominated by Convolutional Neural Networks (CNNs). ResNet-50 is a widely used CNN architecture that is part of the ResNet (Residual Network) family, which was introduced in the paper “Deep Residual Learning for Image Recognition" by He et al. in 2015. The key innovation of ResNet is the introduction of residual connections or skip connections, which help mitigate the vanishing gradient problem that typically hampers the training of deep networks. ResNet-50 is a deep network with 50 layers, including convolutional layers, batch normalization layers, Rectified Linear Unit (ReLU) activation functions, and fully connected layers. ReLU allows only positive values to pass through, introducing non-linearity into the network, which is essential for the network to learn complex patterns in the data. ResNet-50 is a standard choice for many computer vision tasks due to its balance between complexity and performance. The core idea behind ResNet-50 is the residual block, which allows the network to learn residual functions with reference to the input of a layer, instead of learning unreferenced functions. This makes training easier and improves the network's performance. Like many modern architectures, ResNet-50 is often pretrained on large datasets (like ImageNet), which helps the model learn general features before being fine-tuned on specific tasks.

[0059] Swin Transformer is a vision Transformer that capably serves as a general-purpose backbone for computer vision. To address the challenges in adapting the Transformer from language to vision, Swin Transformer presents a hierarchical Transformer with a shifted windowing scheme, which brings greater efficiency by limiting self-attention computation to non-overlapping local windows while also allowing for cross-window connection. The qualities of Swin Transformer make it compatible with a variety of vision tasks, including image classification, object detection and semantic segmentation. Its performance surpasses the previous state-of-the-art, achieving a top 1 accuracy 87.3% on lmageNet-1 K image classification. This demonstrates the potential of Transformer-based models as vision backbones.

[0060] One aspect of the present disclosure relates to T cell morphology during immune synapse formation or T cell spreading. T cell immune synapse is a specialized structure formed between T cells and antigen-presenting cells (APC), characterized by the spatial organization of various molecules into distinct regions. T cell immune synapse formation is driven by the reorganization of cytoskeleton. Upon antigen recognition, adhesion molecules are active and clustered at cell-cell interface, providing anchorage points for actin polymerization. Actin polymerization at the leading edge of cell membrane generates a protrusive force that extends the cell surface, while actin-myosin interactions provide a contractile force to stabilize the spread. The shape of the immune synapse is crucial for regulating and coordinating signaling events, ensuring efficient antigen recognition and overall immune response effectiveness. Accordingly, the present disclosure provides a morphology-based assay to predict T cell long term proliferation. In cases that T cells fail to expand, the expansion can be rescued by adjusting the environmental stiffness, given the previous finding that softer material can enhance T cell proliferation.

[0061] The present disclosure provides the T cell morphological changes influenced by both intrinsic states and extrinsic environment. It’s found that disease states affect T cell spreading. Specifically, T cells from CLL patients are smaller and rounder than those from healthy donors when interacting with functionalized biomaterial, which mimics the immune synapse formation. Turning to the extrinsic factor of extracellular stiffness, it’s found that T cells from both healthy donors and CLL patients exhibited changes in area and roundness as a function of substrate stiffness. However, the mechanosensing effect was more pronounced for cells from healthy donors. The present disclosure also investigated the effects of cytoskeletal protein inhibitors to understand the contributions of different dynamics on cell morphology. Inhibition of actin polymerization using CK666 significantly reduced cell spreading, and inhibition of actomyosin using Y-27632 eliminated the mechaonsensing response. An exploratory PCA including 11 morphological features was performed, and the projection of data on PC1 (78% variance) and PC2 (12% variance) revealed the variance between healthy donors and CLL patients. PCA also identified important features contributing to the variance. To classify the intrinsic state of cells, feature-based machine learning approaches were employed to classify healthy and CLL by inputting morphological features. Random forest model reached relatively higher accuracy (0.753).

[0062] The present disclosure also provides an image-based deep-learning model to classify both disease states and environmental stiffness using T-cell fluorescent images as input. Two architectures, Swin Transformer and ResNet-50, were compared, and Swin Transformer showed superior performance. Pretrained and unpretrained Swin Transformer were also evaluated, and the pretrained model showed better performance, demonstrating the impact of transfer learning on performance. Additionally, within the pretrained Swin Transformer, the effects of freezing and unfreezing the feature extraction weights were compared, and the results showed that unfreezing the weights enhanced the performance. The imagebased deep learning model significantly improved classification accuracy compared to feature-based machine learning model, with the same cohort of T cell images.

[0063] The present disclosure also provides a deep learning-based regression model to predict max doublings using T cell spreading images as input. The ground truth max doublings of ten donors over three stiffness conditions were obtained by proliferation assays. The scatter plot of predicted and true labels showed a monotonic correlation, indicating that T cell short-term spreading images can predict long-term proliferation. In the postprocessing, the predicted max doublings across stiffness were ranked and the best stiffness was selected.

[0064] Another aspect of the present disclosure relates to the mechanosensing response reflected by T cell spreading area across different subsets. The comparison between CD4⁺ and CD8⁺ T cells indicated that CD4⁺ T cells exhibited a dominant mechanosensitive response compared to CD8⁺ T cells. However, this dominance appears to be influenced by the relatively lower percentage of CD4⁺ T cells within the sample. The comparison among four subsets revealed the difference between healthy donors and CLL patients. Specifically, in healthy donors, T cell subsets exhibited a uniform mechanosensitive response, reflecting coordinated behavior across the different T cell types. In contrast, CLL patients demonstrated a more variable mechanosensory response among T cell subsets.

[0065] Accordingly, one embodiment of the present disclosure is a method for rapid assessing the health of a cell from a subject in need thereof. This method comprises the steps of: (a) training a model that provides a standard reference indicating the health of the cell, comprising: (i) isolating the cell from a healthy subject and expanding them on a functionalized substrate; (ii) obtaining raw images for the cell by immunofluorescence and fixed imaging; (iii) processing the raw images to generate sliding patches; (iv) measuring a number of parameters based on the sliding patches; (v) assigning an indicator of health to the cell and recording the measured values for the parameters; (vi) repeating steps (a-i) to (a-v) with the cell from a number of healthy subjects; (vii) consolidating, analyzing and adjusting the measured values for the parameters from all tested healthy subjects to obtain a classifier for the healthy cell; (viii) repeating steps (a-i) to (a-vii) by isolating the cell from a subject with a condition and expanding them on the same functional substrate, wherein the cell from the subject with the condition receives an indicator of the condition and obtains a classifier for the cell having the condition; (ix) repeating step (a-viii) by isolating the cell from a subject with a different condition and expanding them on the same functional substrate; and (x) obtaining the standard reference by consolidating and analyzing the classifiers for the healthy cell and the cell having any condition, (b) assessing the health of the cell from a subject in need thereof, comprising: (i) isolating the cell from the subject in need thereof and obtaining the measured values for the parameters as described in steps (a-i) to (a- iv); (ii) entering the measured values in step (b-i) into the trained model; and (iii) receiving the indicator of health, and (c) selecting a suitable cell expansion plan for the subject.

[0066] Another embodiment of the present disclosure is a method for predicting the optimal growth condition for expansion of a cell from a subject in need thereof. This method comprises the steps of: (a) training a model that provides a standard reference predicting the optimal growth condition for expansion of the cell, comprising: (i) isolating the cell from a healthy subject and expanding them on a functionalized substrate; (ii) obtaining raw images for the cell by immunofluorescence and fixed imaging; (iii) processing the raw images to generate sliding patches; (iv) measuring a number of parameters based on the sliding patches; (v) assigning an expansion indicator to the cell and recording the measured values for the parameters; (vi) repeating steps (a-i) to (a-v) with the cell from a number of healthy subjects; (vii) consolidating, analyzing and adjusting the measured values for the parameters from all tested healthy subjects to obtain an expansion classifier for the healthy cell on the functionalized substrate; (viii) repeating steps (a-i) to (a-vii) by expanding the cell from healthy subjects on a different functionalized substrate to receive a different expansion indicator and obtain an expansion classifier on the different functionalized substrate; (ix) changing the functionalized substrate and repeating steps (a-viii); (x) repeating steps (a-i) to (a-ix) by isolating the cell from a subject with a condition; (xi) repeating step (a-x) by isolating the cell from a subject with a different condition; and (xii) obtaining the standard reference by consolidating and analyzing the classifiers for the healthy cell and the cell having any condition, (b) predicting the optimal growth condition for expansion of the cell from a subject in need thereof, comprising: (i) isolating the cell from the subject in need thereof and obtaining the measured values for the parameters as described in steps (a-i) to (a-iv); (ii) entering the measured values in step (b-i) into the trained model; and (iii) receiving the expansion indicator, and (c) selecting an optimal cell expansion plan for the subject.

[0067] In some embodiments of the present disclosure, the cell assessed and/or expanded by methods disclosed herein is a lymphocyte. Examples of a lymphocyte include, but not limited to, a T cell, a B cell, a natural killer (NK) cell, and a natural killer T (NKT) cell. In some embodiments, the cell is a T cell. In some embodiments, the cell is a natural killer (NK) cell.

[0068] In some embodiments of the present disclosure, the optimal growth condition for the cell expansion includes the stiffness, format, and/or coating of the functionalized substrate. In some embodiments, the functionalized substrate is prepared with poly (dimethyl siloxane) (PDMS), with Young’s modulus in the range of 50 kPa to 2000 kPa, preferably, 250 kPa to 550 kPa. In some embodiments, the PDMS substrate is coated with antibodies.

[0069] In some embodiments of the present disclosure, the model trained in methods disclosed herein can be any deep learning model for image classification that is available such as, e.g., convolutional neural networks (CNNs), AlexNet, ResNet, Vision Transformers (ViTs), and Swin Transformer, or those to be developed in the future.

[0070] In some embodiments of the present disclosure, the parameters are morphology features that describe the characteristics such as, e.g., size, shape, structure, color and pattern of a cell. In some embodiments of the present disclosure, the parameters are selected from the group consisting of area, roundness, major, minor, perimeter, solidity, circularity, height, width, Feret’s diameter, aspect ratio (AR), and combinations thereof.

[0071] As used herein, “area” of a cell refers to the total surface area of a cell's outer membrane, essentially the total space covered by the cell's exterior, calculated by measuring the area of the cell's visible surface when viewed under a microscope.

[0072] As used herein, “roundness” of a cell refers to a quantitative measurement of how close a cell's shape is to a perfect circle, essentially describing how circular a cell appears, with a higher roundness value indicating a more circular shape and a lower value indicating a more elongated or irregular shape.

[0073] As used herein, “major” of a cell refers to the longer axis of an ellipse fit to the shape of the cell. It provides a measure of how long a cell is when approximated by an ellipse.

[0074] As used herein, “minor” of a cell refers to the shorter axis of an ellipse fit to the shape of the cell. The "minor" axis is a measure of how wide a cell is when measured perpendicular to the "major" axis.

[0075] As used herein, “perimeter” of a cell refers to the total distance around the edge of a cell membrane, essentially the length of the cell membrane boundary, which marks the outer limit of the cell itself.

[0076] As used herein, “solidity” of a cell refers to a measure of how compact a cell is, calculated as the ratio of the cell's area to the area of its convex hull; essentially, it indicates how much a cell's shape resembles a perfect circle, with a value closer to 1 signifying a more compact, regular cell and a lower value indicating a more irregular or indented cell shape.

[0077] As used herein, “circularity” of a cell refers to a quantitative measure of how close a cell's shape is to a perfect circle, essentially indicating how round the cell is; a higher circularity value (close to 1) means the cell is more circular, while a lower value signifies a more elongated or irregular shape. It is calculated by dividing the perimeter of a cell by the square root of its area, and is used to analyze cell morphology and identify changes in cell shape during processes like differentiation, migration, or disease progression. [0078] As used herein, “height” of a cell refers to the distance between the top of a cell membrane and the bottom of the cell, often influenced by the size of the nucleus within the cell.

[0079] As used herein, “width” of a cell refers to the distance between one side of a cell membrane and the other side of the cell membrane, like cell height, it is also often influenced by the size of the nucleus within the cell.

[0080] As used herein, “Feret’s diameter” of a cell refers to the distance between two parallel tangents drawn on opposite sides of the cell boundary, perpendicular to the line connecting the two points, essentially representing the maximum "caliper" measurement of the cell along any axis, providing a measure of its size regardless of its shape

[0081] As used herein, “aspect ratio” or “AR” of a cell refers to essentially representing the maximum "caliper" measurement of the cell along any axis, providing a measure of its size regardless of its shape. A higher aspect ratio indicates a more elongated cell, while a lower aspect ratio means a more rounded cell.

[0082] In some embodiments of the present disclosure, the parameters further comprise a proliferation index such as, e.g., a max doubling value. As under herein, a “proliferation index” is a measurement of how many times cells have divided. As used herein, a “max doubling value” refers to the longest period of time it takes for a single cell to divide and produce two daughter cells, or a population of cells to double in quantity.

[0083] In some embodiments of the present disclosure, the proliferation index is obtained by training a regression model using one or more parameters disclosed herein including but not limited to area, roundness, major, minor, perimeter, solidity, circularity, height, width, Feret’s diameter, aspect ratio, and combinations thereof. In some embodiments of the present disclosure, the regression model uses additional parameters. For example, in some embodiments, the parameters further comprise a cell subtype marker and/or a cytoskeletal component.

[0084] As used herein, a “cell subtype marker” is a gene or protein that is specific to a cell subtype. In some embodiments of the present disclosure, the cell subtype marker is selected from the group consisting of T cell receptor (TCR), CD3, CD4, CD8, FoxP3, CD28, CD45RA, CD45RO, CD62L, CCR7, CD27, CD28, CD25, CD127, CD57, CD137; CD19, CD24, CD38, CD40, CD1 D, IgM; CD11 b, CD27, CD161 , CCR7, CD244, NCR3, CD94, CD122, CD69, and combinations thereof.

[0085] As used herein, a “cytoskeletal component” refers to protein filaments that make up the cytoskeleton. In some embodiments of the present disclosure, the cytoskeletal component is selected from microtubules, intermediate filaments, microfilaments, and combinations thereof.

[0086] It is to be understood that the raw images captured and used in methods disclosed herein (e.g., in step (a-ii)) are not limited to static images. For example, in some embodiments of the present disclosure, the raw images can further include those obtained by time-lapse videos. Accordingly, in some embodiments, the parameters used in the methods disclosed herein further comprise temporal features selected from speed and direction of cell movement.

[0087] In some embodiments of the present disclosure, the condition is selected from an autoimmune disease, fibrosis, a viral infection, a transplant rejection, and a cancer. In some embodiments of the present disclosure, the condition is acute lymphoblastic leukemia (ALL) or chronic lymphocytic leukemia (CLL). In some embodiments, the condition is CLL.

[0088] A further embodiment of the present disclosure is a method for treating or ameliorating the effects of a disease in a subject in need thereof. This method comprises: (a) assessing the health of T cells isolated from the subject according to the method disclosed herein; (b) selecting an optimal growth condition for T cells expansion for the subject according to the method disclosed herein; (c) expanding T cells for the subject under the optimal growth condition of step (b) to a clinically relevant number; (d) reinfusing the expanded T cells to the subject; (e) after a course of treatment, reassessing the health of T cells in the subject by repeating step (a); (f) if the reassessment result in step (e) is acceptable, continuing expanding T cells under the current optimal growth condition, or if the reassessment result in step (e) is unacceptable, adjusting T cells expansion by selecting a different optimal growth condition by repeating step (b) and expanding T cells under the adjusted optimal growth condition; (g) reinfusing the expanded T cells in step (f) to the subject; and (h) repeating steps (e) to (g) as necessary. [0089] As used herein, the terms "treat," "treating," "treatment" and grammatical variations thereof mean subjecting an individual subject to a protocol, regimen, process or remedy, in which it is desired to obtain a physiologic response or outcome in that subject, e.g., a patient. In particular, the methods and compositions of the present disclosure may be used to slow the development of disease symptoms or delay the onset of the disease or condition, or halt the progression of disease development. However, because every treated subject may not respond to a particular treatment protocol, regimen, process or remedy, treating does not require that the desired physiologic response or outcome be achieved in each and every subject or subject population, e.g., patient population. Accordingly, a given subject or subject population, e.g., patient population, may fail to respond or respond inadequately to treatment.

[0090] As used herein, the terms “ameliorate”, "ameliorating" and grammatical variations thereof mean to decrease the severity of the symptoms of a disease in a subject.

[0091] As used herein, a “subject” is a mammal, preferably, a human. In addition to humans, categories of mammals within the scope of the present disclosure include, for example, agricultural animals, veterinary animals, laboratory animals, etc. Some examples of agricultural animals include cows, pigs, horses, goats, etc. Some examples of veterinary animals include dogs, cats, etc. Some examples of laboratory animals include primates, rats, mice, rabbits, guinea pigs, etc. In the context of the present disclosure, the phrase “a subject in need thereof” means a subject receiving a treatment with the cells assessed and expanded using methods disclosed herein.

[0092] In some embodiments of the present disclosure, the disease is selected from the group consisting of an autoimmune disease, fibrosis, a viral infection, a transplant rejection, a cancer, or combinations thereof.

[0093] In some embodiments of the present disclosure, the autoimmune disease is selected from the group consisting of type 1 diabetes, systemic lupus erythematosus, Sjogren’s syndrome, diffuse scleroderma, inflammatory myopathy, ANCA-associated systemic vasculitis, antiphospholipid syndrome, mucosal-dominant pemphigus vulgaris, anti-MuSK-antibody-positive myasthenia gravis, generalised myasthenia gravis, lupus nephritis, neuromyelitis optica spectrum disorder, myasthenia gravis, chronic inflammatory demyelinating, polyradiculoneuropathy, immune-mediated necrotising myopathy, immune nephritis, refractory POEMS syndrome, amyloidosis, autoimmune haemolytic anaemia, vasculitis, Crohn’s disease, ulcerative Colitis, dermatomyositis, and Still disease.

[0094] In some embodiments of the present disclosure, the viral infection is caused by a virus selected from human immunodeficiency virus (HIV), hepatitis B virus (HBV), hepatitis C virus (HCV), and cytomegalovirus (CMV).

[0095] In some embodiments of the present disclosure, the transplant rejection is selected from HLA-A2 mismatched liver transplantation, and HLA-A2 mismatched living donor kidney transplantation.

[0096] In some embodiments of the present disclosure, the cancer is a hematologic cancer selected from the group consisting of acute lymphoblastic leukemia (ALL), chronic lymphocytic leukemia (CLL), follicular lymphoma, mantle cell lymphoma, diffuse large B-cell lymphoma, and multiple myeloma.

[0097] In some embodiments of the present disclosure, the cancer is a solid tumor selected from the group consisting of glioblastoma, ependymoma, medulloblastoma, pediatric brain tumor, breast cancer, neuroblastoma, liver cancer, pancreatic cancer, prostate cancer, lung cancer, gastric cancer, and esophageal cancer.

[0098] In some embodiments of the present disclosure, the disease is chronic lymphocytic leukemia (CLL).

[0099] In some embodiments of the present disclosure, the subject is a mammal. In some embodiments, the subject is a human.

[00100] In some embodiments of the present disclosure, the cell used for treatment is a lymphocyte other than a T cell. For example, the cell can be a B cell, a natural killer (NK) cell, or a natural killer T (NKT) cell. In some embodiments, the cell used for treatment is a natural killer (NK) cell or a natural killer T (NKT) cell.

[0100] In some embodiments of the present disclosure, the treatment method further comprises administering to the subject one or more standard therapy for the disease. Thus, the assessment and/or adjustment steps in the method disclosed herein can also be used to assess and/or mitigate any impact of the other therapies on the cells of the subject.

[0101] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. As used in the specification and the appended claims, the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise.

[0102] The following examples are provided to further illustrate the methods of the present disclosure. These examples are illustrative only and are not intended to limit the scope of the disclosure in any way.

EXAMPLES

Example 1

Assaying and Classifying T Cell Function by Cell Morphology

[0103] Cells of the immune system execute and coordinate a wide range of functions in both normal and pathological physiologies. These cells are also a compelling platform for targeted, effective, and persistent therapy for a range of diseases, seen most prominently in the successful clinical deployment of T cells against cancer. However, the use of cells as a “living drug” poses several challenges, including wide variability in functionality between individuals as a result of disease state. In the case of T cell therapy against cancer, long-term diseases such as Chronic Lymphocytic Leukemia (CLL) induce T cell deficiencies resembling cellular exhaustion, complicating the preparation of therapeutic quantities of cells, and ensuring efficacy once reintroduced to the patient. The ability to rapidly estimate the responsiveness of an individual’s T cells would dramatically improve cell production by tailoring ex vivo culture conditions to an individual’s starting material and provide powerful insight into T cell health over the course of treatment. Current assays in this direction include cell count, biomarkers, and cytokine secretion. However, our group demonstrated that assays of high-level cellular function, specifically cell migration, provide enhanced insight into long-term cell function compared to these molecular measures alone (Figures 1A-1E). [0104] This Example investigated cell morphology as a simpler indicator of T cell functionality, reflecting both the intrinsic state of an individual’s cells as well as response to the surrounding environment. In these assays, T cells are allowed to interact with planar test surfaces under controlled ex vivo conditions. Like other cells, T cells undergo a phase of rapid spreading, driven by actin polymerization, followed by contraction of this cytoskeletal network. The shape of cells on these surfaces reflects a balance of intracellular processes and the interaction of cells with the extracellular environment. Consequently, measures of cell spreading, such as area, have been used as surrogates of T cell activation and subsequent function. This Example refined this basic approach to capture changes in cell morphology as a function of two different types of factors. As a cell-intrinsic factor, T cells isolated from healthy donors were compared against counterparts from patients being treated for chronic lymphocytic leukemia (CLL), a disease often accompanied by T cell exhaustion. As a complementary, extrinsic factor, cell response to substrates of different mechanical stiffness was examined as well. This was inspired by a growing body of knowledge that T cells respond to the mechanical resistance of their environment, altering a range of readouts from cytokine secretion to long-term proliferation. This Example examined how morphological outputs can be associated with these two factors, providing a measure of the overall cellular states.

Materials and Methods

Glass-Supported Poly (dimethyl siloxane) (PDMS) Substrate Preparation

[0105] PDMS substrates of varying stiffness were prepared following established protocols. Blending Sylgard 527 and Sylgard 184 in mass ratios of 10:1 , 3:1 , and 1 :3 produced substrates with Young’s modulus of 250 kPa, 1000 kPa, and 2000 kPa. Thin PDMS layers were created on glass coverslips (thickness #0, Electron Microscopy Sciences, Hatfield, PA, USA). A droplet of PDMS mixture was pressed using a PDMS cube to flatten the droplet and create a thin (~20 pm) layer, which would be peeled off after curing overnight at 65 °C. The PDMS cubes were silanized overnight with (tridecafluoro-1 ,1 ,2, 2,-tetrahydrooctyl)-1 -trichlorosilane (United Chemical Technologies, Bristol, PA, USA), to facilitate removal from the PDMS substrates. To activate T cells, each PDMS substrate was coated overnight at 4 °C with a mixture of a-CD3 (clone OKT3, Bio X Cell, Lebanon, New Hampshir, USA) and a-CD28 (clone 9.3, Bio X Cell, Lebanon, New Hampshir, USA) antibodies in a mass ratio of 1 :1 for a total concentration of 20 pg/mL in PBS. For visualization and quantification of surface-bound protein, half the antibodies in this mix were labeled with Alexa Fluor 568 NHS Ester (Succinimidyl Ester) (Thermo Fisher Scientific, Frederick, MD, USA).

PDMS Substrate Stiffness Characterization

[0106] Young’s modulus (E) of prepared PDMS substrates was measured by indentation. Thick slabs (several mm) of PDMS were deformed using a flat cylindrical head with a calibrated mass. The material’s Young’s modulus was estimated from the head diameter (D, 12 mm), deflection (h), weight (m), gravitational constant (g), and Poisson ratio (v) of 0.5 assuming Hertzian contact with the following equation:

[0107] The thickness of PDMS film on glass coverslip was measured by microscopy. The surface of glass coverslip was labeled with marker and the surface of PDMS film was coated with fluorescently labeled antibody. The z values of both glass coverslip and PDMS surface was noted down from imaging, and the thickness was calculated by subtraction, which is about 20 pm.

Cell Isolation and Culture

[0108] Mixed CD4⁺/CD8⁺ primary human T cells were isolated from Leukapheresis packs derived from healthy adult donors (New York Blood Center) and CLL patients (Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA, USA), using negative selection (RosetteSep kit, Stem Cell Technology, Vancouver, BC, Canada) and gradient centrifugation (Ficoll-Paque Premium, Cytiva, Uppsala, Sweden); cells were not purified on the basis of subtype, and consequently these preparations contained a mix of naive, memory, and effector phenotypes. Cells were cultured in complete culture media consisting of RPMI 1640 medium (Gibco, Grand Island, NY, USA) supplemented with 10 mM HEPES (Gibco, Grand Island, NY, USA), 10 mM L-Glutamine (Gibco, Grand Island, NY, USA), 10% (v/v) fetal bovine serum (FBS; Gibco, Grand Island, NY, USA), 0.34% (v/v) 0- mercaptoethanol (Sigma-Aldrich, Burlington, Massachusetts), and 10 mM penicillinstreptomycin (Gibco, Grand Island, NY, USA). After isolation, cells were frozen in complete media with 40% FBS and 10% DMSO in liquid nitrogen. Before experiments, cells were thawed and rested under standard culture conditions (37 °C, 5% 002/95% air) overnight.

Assays of Cell Spreading

[0109] T cells were seeded onto glass-supported PDMS substrates at a concentration of 1 x 10⁶ cells/mL. Following 40 min T cell spreading, samples were fixed in 4% PFA for 20 min at room temperature and permeabilized with 0.1 % Triton X for 10 min at room temperature. Then, samples were stained with Alexa Fluor 488 phalloidin (Thermo Fisher Scientific, Frederick, MD, USA) at 1 :40 dilution for 20 min at room temperature, followed by washing twice. Samples were then imaged using an Olympus 1X81 inverted microscope, equipped with an Andor iXon EMCCD camera, providing a 1002 x 1002 array of 8 pm x 8 pm pixels. Live imaging was conducted by live-cell microscopy under 60x magnification and bright field in the first 60 min after seeding T cells onto the PDMS substrate, using a stage top incubator (Tokai Hit, Bala Cynwyd, PA, USA). The image was collected at 30-s intervals over the 60 min observation period. For analysis in live imaging, within the 120 frames, 2- 5 cells were tracked and measured every 10 frames (5 min) by manual segmentation and measuring cell area on Imaged.

[0110] Fixed imaging was performed under 40x. Image analysis was performed in Imaged using functions of Smoothing, Thresholding, Set Measurement, and Analyze Particles to measure morphological features of single cells, including Area, Perimeter, Width, Height, Major, Minor, Circularity, Feret’s Diameter, Aspect Ratio (AR), Roundness, and Solidity. Only accurately segmented single cells were selected by quality control for further analysis.

Flow Cytometry

[0111] Flow cytometry to check the purity of isolated T cells was performed on a FACSCanto II (BD Biosciences, Franklin Lakes, Nd, USA) with a minimum of 10,000 gated events. Analysis was performed on FCS Express V6 (De Novo Software, Pasadena, CA, USA).

Inhibitor Studies [0112] Arp2/3 complex inhibitor CK666 (Sigma-Aldrich, Burlington, Massachusetts) (100 pM) was used to inhibit actin polymerization. ROCK inhibitor Y- 27632 (Sigma-Aldrich, Burlington, Massachusetts) (60 pM) was used to inhibit actomyosin contractility. Cells were pretreated with either CK666 or Y-27632 in complete culture media at 37 °C for 15 min and then were seeded onto the prepared PDMS substrates. The cells spread in the presence of the inhibitor for 40 min, followed by fixation, permeabilization, and incubation with Alexa Fluor 488 phalloidin, as previously described.

Statistical Analysis

[0113] llnpair t-tests with Welch’s correction and two-way ANOVA with Tukey multiple comparison test were conducted on GraphPad Prism 9.4.0 for quantitative comparisons of T cell Area and Roundness between Healthy and CLL, and across different PDMS stiffness conditions. Each data point in the bar graphs represents the average for that metric of all cells (approximately 100 cells) on a single PDMS sample, while data for all individual cells were included in analysis. Three healthy donors (H1 , H2 and H3) and six CLL patients (D2, D18, D51 , D63, D 67, D75) were included. For H1 , n=4. For H2, n = 8. For H3, n = 6. For CLL patients, n = 2 due to the limited availability of T cells in samples.

Data Preprocessing and Data Normalization

[0114] Single-cell morphological features were acquired from spreading assays. These features include Area, Perimeter, Width, Height, Major, Minor, Circularity, Feret’s Diameter, Aspect Ratio (AR), Roundness, and Solidity. Only accurately segmented single cells were selected by manually labeling the segmentation quality with “good” or “bad.” There are 12,101 accurately segmented single cells in total. The dataset was then aggregated into sample level - 100 samples in total. The values of morphological features were normalized using algorithm preprocessing. StandardScaler from scikit-learn (sklearn) library.

Principal Component Analysis (PCA)

[0115] Principal Component Analysis (PCA) was performed on the normalized morphological feature dataset using the ‘PC class from the ‘sklearn. decomposition’ module. The first two principal components, which explain the majority of the variance, were retained to plot data. Feature-Based Classification

[0116] Binary labels were assigned to Healthy (0) and CLL (1) samples. There are 45 CLL (1) samples and 55 Healthy (0) samples. The study employed three distinct classification models, namely Single-Feature Decision Tree, Multi-Feature Decision Tree, and Random Forest, to classify samples into Healthy or CLL. Decision Tree was performed using the algorithm DecisionTreeClassifier from the sklearn library. For Single-Feature Decision Tree, only the feature Area was input for classification because Area was the primary feature focused on in statistical analysis to compare Healthy and CLL. Similarly, Random Forest was performed using the algorithm Random Forestclassifier from the sklearn library. Default hyperparameters were used as provided by the sklearn library to ensure reproducibility of our results across different studies. The models were compared through the average performance of three independent runs of 10-fold cross-validation. To prevent data leakage, each fold involved the random selection of one healthy donor and two CLL patients as the testing dataset. The evaluation metrics, including Accuracy, Area Under Curve (AUC), Sensitivity, Specificity, and Mathew’s correlation coefficient (MCC) were calculated to assess and compare the performance of the models.

Results

Glass-supported Poly (dimethyl siloxane) (PDMS) substrate preparation and assay optimization

[0117] PDMS with varied stiffness was fabricated by blending commercially available Sylgard 527 and Sylgard 184 with different ratios (Figure 2A). PDMS thin films were cured on top of glass coverslips (Figure 2B) and were coated with anti- CD3 and anti-CD28 antibody mixture. The ratio of anti-CD3:anti-CD28 was optimized, and the ratio of 1 :1 was selected to reflect the most significant mechanosensitive spreading (Figure 2C). The antibodies were conjugated with fluorescent dye for visualization (Figure 3A) and quantification to make sure the amount of activating antibody across stiffness is the same (Figure 3B). The functionalized PDMS substrate was validated by seeding the cells. T cells were able to attach and spread on the functionalized PDMS substrate, while they were not able to spread on the negative control, PDMS without activating the antibody (Figure 3C). [0118] Cells were allowed to spread for 40 min followed by fixing, permeabilization, and staining (Figure 4). The spreading time was optimized by tracking the dynamics of T cell spreading and comparing multiple time points. Live imaging captured T cell spreading over 60 min and identified 40 min as a time point during which cells stabilize on each surface and produce the greatest difference as a function of elastic modulus (Figure 5A). Fixed imaging analysis comparing 20 min, 30 min, and 40 min timepoints also validated that the 40 min timepoint retains the stability and resolution seen in the live-cell assays (Figure 5B). Fixed imaging under 40x magnification throughout the study was chosen because it balances well between cell number acquired in a field of view and the measurement precision after comparing with 100x magnification (Figures 6A-6C).

Disease State Affects T Cell Morphology

[0119] Primary human T cells from 3 healthy (H) donors and 6 CLL patients (Table 1.1) were allowed to spread on polydimethylsiloxane (PDMS) surfaces coated with a-CD3/CD28 (Figure 7A). These substrates were prepared by mixing two standard formulations of PDMS, Sylgard 527 and Sylgard 184 (Dow Silicones Corporation, Midland, Ml, USA). Elastic (Young’s) modulus was modulated by changing the ratio of the two formulations, producing three different stiffnesses of 250 (Soft, 10:1 ratio of 527:184), 1000 (Medium, 3:1 ), and 2000 (Hard, 1 :3) kPa PDMS (Figure 7B). Comparison of the fluorescence intensity of Alexa 586-labeled OKT3/9.3 showed that the concentration of adsorbed antibodies was similar across the three different formulations (Figure 7C). Initial live-cell experiments captured the dynamics of cell spreading on these surfaces, identifying 40 min as a timepoint during which cell area, a representative measure of this interaction, stabilized on each surface and also produced the greatest difference as a function of elastic modulus. Toward a readily deployable assay of cell function, it was focused on samples that were fixed at specific time points; the 40 min timepoint retained the stability and resolution seen in the live-cell assays and was chosen as a standard timepoint for the remainder of this study.

Table 1.1 : CLL Patient Information.

[0120] Representative images of fixed T cells from healthy donors and CLL patients on surfaces of different elastic modulus are shown in Figure 7D. Having been purified using techniques that are independent of subtype, these samples contained a mix of naive, memory, and effector cells that are representative of the donor population. Most prominently, cells from CLL patients appear smaller than those from the healthy counterparts, which applies to all three substrates stiffness (Figure 8A). Notably, cells from CLL patients showed higher Roundness than cells from healthy donors, supporting the concept that donor disease state affects cell morphology.

T Cells from Both Healthy Donors and CLL Patients Respond to Substrate Stiffness in Spreading, with more pronounced effect for healthy donors

[0121] Turning to the extrinsic factor of extracellular stiffness, T cells from both healthy and CLL donors exhibited changes in Area and Roundness as a function of substrate modulus (Figure 8B). Donor-to-donor variation existed in terms of sensing different stiffness (Figures 9A-9B). Notably, the mechanosensing effect was more pronounced for cells from healthy donors than CLL patients, suggesting a functional impact of disease state and exhaustion on this response. Observation and analysis from live imaging also resonated with this finding (Figures 8B, 9A-9B, and 10A- 10B).

[0122] While these experiments demonstrated an impact of disease state and substrate stiffness on two morphological features (Area and Roundness), the goal is to find a way to identify the experimental parameters associated with a sample from cell morphology. The differing responses of Area and Roundness on disease state and substrate stiffness make the use of either feature individually complicated for this purpose. As such, the next sections expanded this approach to include additional morphological features.

RCA Reveals the Variance between CLL and Healthy T Cells and Identifies Important Morphological Features Contributing to the Variance

[0123] Fixed imaging was performed under both 40x. Image analysis was performed in ImageJ using functions of Smoothing, Thresholding, Set Measurement, and Analyze Particles to measure morphological features of single cells, including Area, Perimeter, Width, Height, Major, Minor, Circularity, Feret’s Diameter, Aspect Ratio (AR), Roundness, and Solidity. Only accurately segmented single cells were selected by quality control for further analysis (Figure 11).

[0124] A dataset containing 11 morphological features was generated from T cells as a function of disease state and substrate stiffness. Dimensionality reduction of this data by Principal Component Analysis (PCA) revealed a first component (PC1) that strongly expressed the dataset, accounting for 78% of the variance; PC2 was associated with 12%, leaving 10% to the remaining principal components (Figure 12A). Examination of PC1 identified a set of key features with similar and large weights, including Area, Feret’s Diameter, and Minor (Figure 12B). These weights also captured the anti -correlation between Area and Roundness suggested in Figures 8A-8B. However, the lower weight of Roundness in PC1 does suggest that these two parameters are not simply anti-correlated, and they may provide additional information useful in identifying the intrinsic and extrinsic conditions. Notably, the projection of the data along PC1 showed the separation of datasets based on disease state (Figure 12A, red vs. blue).

Machine Learning Classifies CLL Patients Based on Morphological Features

[0125] While PCA analysis showed the promise of using multiple features to distinguish between intrinsic and extrinsic parameters, there are several limitations to this approach. Most prominently, the identification of cell spreading conditions from morphological data is a classification rather than an analysis question. In addition, PCA is fundamentally a linear analysis approach, while the relationship between morphological inputs and the experimental conditions is likely more complex. Consequently, this section applies machine learning tools to the classification of cells, focusing first on the disease state. Three different classification models were evaluated and compared through the average performance of three independent runs of 10-fold cross-validation (Table 1 .2).

Table 1.2: Only morphological features as input to classify Healthy vs. CLL [0126] The first model used a Decision Tree with a single feature (Area) as input, resulting in an accuracy of 0.677, an Area Under the Curve (AUC) of 0.683, and a Mathew’s correlation coefficient (MCC) of 0.372. The second model utilized a Decision Tree with 11 morphological features as input, achieving an accuracy of 0.651 , an AUC of 0.650, and a significantly improved MCC of 0.649. The third model employed a Random Forest with all 11 morphological features as input, demonstrating an improved accuracy of 0.753, a notably higher AUC of 0.812, and an MCC of 0.596. This comparison indicates that incorporating multiple morphological features plays an important role in classifying T-cell disease states.

[0127] Furthermore, the impact of including substrate stiffness as an additional input along with the 11 morphological features was investigated to improve the classification accuracy. Two methods of labeling stiffness were employed, including one-hot encoding and normalized Young’s modulus (Tables 1.3 and 1.4). Incorporating stiffness as an additional input led to an improvement across all models, with one hot encoding providing similar performance to the inclusion of modulus as a numeric variable.

Table 1.3: Stiffness as an additional input feature (one hot encoding) to classify Healthy vs. CLL

Table 1.4: Stiffness as an additional input feature (normalized Young’s Modulus) to classify Healthy vs. CLL

Effect of Cytoskeletal Protein Inhibitors on T Cell Response to Substrate Stiffness

[0128] The effects of cytoskeletal protein inhibitors were also investigated to understand the contributions of different dynamics on cell morphology. Inhibition of Arp2/3-based actin branching using CK666 (100 pM) significantly reduced cell spreading. Inhibition of Rho-modulated actomyosin contraction using Y-27632 (60 pM) eliminated the mechanosensing response (Figures 13A-13B). These results show that different types of cytoskeletal dynamics and their associated signaling pathways are represented by the complex measurement of cell morphology.

Effect of I CAM-1 on T Cell Response to Substrate Stiffness

[0129] T cell mechanosensing was investigated in the presence or absence of ICAM-1 , a ligand for the T cell integrin LFA-1 . The result showed that adding ICAM-1 increased T cell area overall, which is consistent with its role as an adhesion molecule (Figure 14B). When T cells bind to ICAM-1 via LFA-1 , it stabilizes the cytoskeletal rearrangements necessary for spreading, promoting more extensive spreading as the cells can form more stable and stronger adhesion points.

[0130] It was also found that in the presence of ICAM-1 did not affect how T cells sense stiffness in spreading. This suggests that the mechanosensing capability of T cells might primarily depend on their intrinsic signaling pathways and cytoskeletal dynamics, which can be triggered by stiffness independent of ICAM-1 engagement (Figure 14A).

T Cell IL-2 Secretion Shows Sensitivity to Substrate Stiffness [0131] To investigate T cell downstream activation function in response to stiffness, IL-2 secretion was measured across three different stiffnesses. First, 3hr and 4hr IL-2 secretion was compared to better understand the kinetics and determine the proper time point for measurement. The result showed that 3hr IL-2 secretion is lower than 4hr, suggesting that the IL-2 secretion is still ongoing at 3hr and the early time point might not provide a clear picture of how substrate stiffness influences T cell activation. By 4hr, T cells have had sufficient time to complete the secretion of IL-2, leading to an overall higher MFI in our flow cytometry data. This time point also reveals how stiffness affects T cell activation (Figure 15A).

[0132] The fact that mechanosensing was evident at 4hr but not at 3hr also suggests that the process of mechanosensing readout in IL-2 secretion requires a more prolonged period of activation to be fully observable. The 4hr time point likely represents a more complete activation profile, where the effects of substrate stiffness on IL-2 secretion can be accurately assessed. Therefore, 4hr was chosen as the time point for the future IL-2 secretion study.

[0133] To further investigate the mechanism of mechanosensitive IL-2 secretion, T cells were treated with 0.16ul Nocodazole (NZ) and Taxol (TX), respectively, in 80ul cell solution, reaching 6.6 uM NZ and 1 uM TX. NZ disrupts microtubules by preventing their polymerization, and TX stabilizes microtubules and prevents their depolymerization. Microtubules are part of the cytoskeleton important for maintaining cell shape, contributing to signal transduction, and are involved in the formation of the immunological synapse and T cell activation.

[0134] The result showed that NZ significantly decreased the percentage of cells secreting IL-2 and the overall secretion level, while TX slightly decreased the percentage of cells secreting IL-2, but it impaired the overall secretion level (Figure 15B). Both of the drugs abrogated the mechanosensitive IL-2 secretion. This result indicates that while TX might initially seem to enhance microtubule function, excessive stabilization can also impair the dynamic reorganization of the cytoskeleton that is necessary for proper signaling. The disappearance of mechanosensing under both treatments suggests that microtubule dynamics are critical for T cells to sense and respond to mechanical cues. Disruption of these dynamics - either by preventing microtubule formation or by excessively stabilizing them - impairs the ability of T cells to adjust their behavior based on environmental stiffness.

Discussion

[0135] An individual’s immune response is a complicated result of multiple factors, including genetics, environment, disease, and lifestyle. This variability impacts cellular therapies based on immune cells, from the success of ex-vivo cell production to the specification of systems for activating immunity in situ. The ability to assess T cell functional response would be transformative to these therapies. The driving concept behind this study is that measures of complex cellular functions such as morphology provide insight into the state of immune cells to a degree not attainable through biomarkers and -omics based technologies; it was previously showed that measures of cell migration provide a better predictor of subsequent function than biomarkers or clinical diagnoses.

[0136] This Example provides a more rapid and deployable approach to describing T cell function, focusing on cell morphology. Most directly, it was shown that donor disease state and cell response to substrate stiffness influence cell morphology. Conversely, it was shown that machine learning approaches combining multiple quantitative measures of morphology have promise in identifying the impact of disease state on an individual’s T cells. Applied to the clinical setting, this approach promises a measure of how exhausted or CLL-like an individual’s cells are following therapy, which could guide subsequent treatment.

[0137] Machine learning-based morphology analysis was less effective in identifying what stiffness of material was used to stimulate the cells. Continued refinement of this model could allow specification of biomaterial properties that optimize ex vivo or in situ activation of T cells, avoiding time- and resourceconsuming trial-and-error approaches. Notably, the analysis workflow used all 11 measures of cell morphology that were collected. Many of these features — such as Major, Minor, Aspect Ratio, and Feret’s Diameter — seem to capture similar aspects of cell spreading. It is tempting to remove measures that have some correlation from the analysis to improve accuracy. However, each of these measures has a specific definition that is distinct from the others. While not fully independent, the inclusion of all parameters to the machine learning workflow has the best opportunity to optimize performance, given sufficient data. Conversely, the inclusion of other measures of morphology that capture features very different from the existing list may improve performance. Finally, it is anticipated that further development of these methods, such as using image-based deep-learning tools, may improve the performance of this approach.

Example 2 Image-Based Deep Learning Outperforms Morphometric Analysis in Classification

[0138] Cells are emerging as powerful “living drugs”, offering new therapeutic possibilities. However, they present unique challenges due to inherent differences between cells from different individuals, resulting in variable long-term outcomes such as tumor-targeting efficacy and success in pre-treatment expansion. Developing a method to predict these outcomes would significantly advance the field. Consequently, various biomarkers, including exhaustion marker expression, cytokine secretion, and phenotypes, are being investigated.

[0139] Recently, it was demonstrated that incorporating whole-cell functionalities, such as migration, alongside traditional biomarkers enhance the accuracy of long-term function prediction. The goal of this Example is to achieve the short-time assay of migration in a simplified format suitable for clinical laboratories. A T cell spreading assay was developed, which showed proof-of-principle that the morphological features measured from T cell spreading images can classify T cell disease state and environmental stiffness, but its performance was limited. Therefore, this Example aimed to leverage an image-based deep learning approach to improve classification accuracy and overall performance.

Materials and Methods

Glass-Supported Poly (dimethyl siloxane) (PDMS) Substrate Preparation

[0140] PDMS substrates of varying stiffness were prepared by tuning the crosslinker to base ratio. Mass ratios of 1 :10 and 1 :50 (crosslinker: base) produced substrates with Young’s modulus of 2300 kPa (hard) and 50 kPa (soft). For intrinsic state classification, only hard surface was used. In the preparation, a droplet of PDMS on glass coverslips (thickness #0, Electron Microscopy Sciences, Hatfield, PA, USA) created thin layer after curing at 65 °C for 16 hours.

[0141] To activate T cells, each PDMS substrate was coated with Purified Goat anti-mouse IgG (clone: Poly4053, Biolegend) (3ug/ml for hard surface and 1 ug/ml for soft surface to matche the 2^nd layer antibody amount, diluted in PBS) at room temperature on shaker for 2 hours, washed with PBS three times, and then coated with a mixture of a-CD3 (clone OKT3, Bio X Cell, Lebanon, New Hampshir, USA) and O-CD28 (clone 9.3, Bio X Cell, Lebanon, New Hampshir, USA) antibodies in a mass ratio of 1:4 for a total concentration of 50 pg/mL in BSA. 10Oul coating solution was added to each PDMS substrate to cover the whole surface.

PDMS Substrate Stiffness Characterization

[0142] Young’s modulus (E) of prepared PDMS substrates was measured by Instron 5848 MicroTester. Thick slabs (1 cm) of PDMS were cured in 35mm- diameter cell culture dish and were mounted onto Instron stage. An indentation head (d=12mm) was mounted. Three cycles of loading 10s and holding 20s were repeated. Specifically, Block 1 : indent 0.4 mm, speed 0.04mm/s; Block 2: hold 20s; Block 3: indent 0.4 mm, speed 0.04mm/s; Block 4: hold 20s; Block 5: indent 0.4 mm, speed 0.04mm/s; Block 6: hold 20s. Load over time was shown in Figure 16A. For analysis, the stable point after holding was found, and displacement over weights was plotted (Figure 16B) to find the slope by linear regression. The weights were calculated from corresponding load (kN). The material’s Young’s modulus was estimated from the head diameter (D, 12 mm), deflection (h), weight (m), gravitational constant (g), and Poisson ratio (v) of 0.5 assuming Hertzian contact with the following equation:

Cell Isolation and Culture

[0143] Mixed CD4⁺/CD8⁺ primary human T cells from 3 healthy adult donors were isolated from Leukapheresis packs (New York Blood Center) by using negative selection (RosetteSep kit, Stem Cell Technology, Vancouver, BC, Canada) and gradient centrifugation (Ficoll-Paque Premium, Cytiva, Uppsala, Sweden). Mixed CD4⁺/CD8⁺ primary human T cells from 7 CLL patients were isolated from peripheral blood mononuclear cells (PBMCs) (Columbia University Irving Medical Center, New York, NY, USA) by using negative selection (Pan-T cell isolation kit, Miltenyi Biotec, Germany). The donor index used in this study can be found in Table 3.1 . Cells were not purified on the basis of subtype, and consequently these preparations contained a mix of naive, memory, and effector phenotypes. Cells were cultured in complete culture media consisting of RPMI 1640 medium (Gibco, Grand Island, NY, USA) supplemented with 10 mM HEPES (Gibco, Grand Island, NY, USA), 10 mM L- Glutamine (Gibco, Grand Island, NY, USA), 10% (v/v) fetal bovine serum (FBS; Gibco, Grand Island, NY, USA), 0.34% (v/v) p-mercaptoethanol (Sigma-Aldrich, Burlington, Massachusetts), and 10 mM penicillin-streptomycin (Gibco, Grand Island, NY, USA). After isolation, cells were frozen in complete media with 40% FBS and 10% DMSO in liquid nitrogen. Before experiments, cells were thawed and rested under standard culture conditions (37 °C, 5% CO2/95% air) overnight.

Assays of Cell Spreading

[0144] T cells were seeded onto glass-supported PDMS substrates at a concentration of 1 x 10⁶ cells/mL. Following 40 min T cell spreading, samples were fixed in 4% PFA for 20 min at room temperature and permeabilized with 0.1 % Triton X for 10 min at room temperature. Then, samples were stained with Alexa Fluor 488 phalloidin (Thermo Fisher Scientific, Frederick, MD, USA) at 1 :40 dilution for 20 min at room temperature, followed by washing twice. Samples were then imaged using an Olympus 1X81 inverted microscope, equipped with an Andor iXon EMCCD camera, providing a 1002 x 1002 array of 8 pm x 8 pm pixels. Fixed imaging was performed under 40x magnification to acquire more cells in a field of view for analysis.

Feature-Based Machine Learning for Classification

[0145] Single-cell morphological features were acquired from T cell images using the established morphometric analysis (Example 1) for this new cohort of donors (3 healthy donors and 7 CLL patients). These features include Area, Perimeter, Width, Height, Major, Minor, Circularity, Feret’s Diameter, Aspect Ratio (AR), Roundness, and Solidity. Only accurately segmented single cells were selected for analysis. A random forest model was developed to automatically label the segmentation outcome with bad (0) or good (1) based on 11 morphological features. The dataset was then aggregated into image level, and image numbers were balanced between two categories with 167 images in Healthy and 155 images in CLL. The values of morphological features were normalized using the algorithm “preprocessing. StandardScaler” from the scikit-learn (sklearn) library. Three distinct classification models, namely Single-Feature Decision Tree, Multi- Feature Decision Tree, and Random Forest were employed to classify samples. Default hyperparameters were used as provided by the sklearn library to ensure reproducibility of our results across different studies. The models were compared through the average performance of three independent runs of 10-fold cross- validation. The evaluation metrics, including Accuracy, Area Under Curve (AUC), Sensitivity, and Specificity were calculated to assess and compare the performance of the models.

Image Processing for Deep Learning

[0146] Fluorescent images were saved and renamed using the format “Donorlndex_ExperimentDate_Stiffness_SpreadingTime_Samplelndex_lmagelndex” . There are 915 images of CLL patients and 462 images of Healthy donors. The raw images were uploaded to the server by FileZilla and processed on Visual Studio Code. Each raw image (1002x1004) was sliced into 25 small patches (224x224), 5 rows x 5 columns. For the 5^th row and 5^th column, the end of the raw image was used as the end of small patches to back draw the patch. The image numbers between two classification categories were balanced. The balanced images patches of both categories were then split into Train, Validation, and Test datasets at a ratio of 80%: 10%: 10%.

Swin Transformer Model for Image Classification

[0147] For classification between Healthy and CLL, Healthy patches and CLL patches were labeled with 0 and 1. To balance image numbers between Healthy (0) and CLL (1), half of the patches from the CLL folder were randomly selected for the classification task. Train, Validation, and Test datasets were loaded using torch. utils.data.DataLoader(). A pre-trained SWIN-Transformer or ResNet-50 was loaded for feature extraction, and the classification head was modified appropriately for our binary classification purposes. In the new head, there is a linear layer with ‘n_inputs=768” input features and 64 output features, a ReLU activation function, a dropout layer with a 20% dropout rate, followed by a linear layer with 64 input features and 1 output feature, and lastly a Sigmoid activation function.

[0148] This model was trained with image patches as input, using a binary cross entropy (BCE) loss function. The model was trained for 100 epochs. After patch-level classification, a majority voting mechanism was incorporated for imagelevel classification. If more than 50% of the patches derived from an image were predicted as 0/1 , then the image was predicted as the corresponding label. Classification accuracy, AUC, sensitivity, specificity, and confusion matrix were calculated to measure the model performance.

[0149] Pretrained and unpretrained Swin Transformer model was compared by setting “pretrained=True” or “pretrained=False”. The comparison was also conducted between freezing and unfreezing the feature extraction weights by setting “param. requires_grad=False” or “param. requires_grad=True”.

[0150] For classification between Hard and Soft surfaces, cell patches from Soft and Hard substrates were labeled with 0 and 1 . The train dataset has 9620 Hard patches and 9040 Soft patches; the validation dataset has 1202 Hard patches and 1130 Soft patches; the test dataset has 1203 Hard patches and 1130 Soft patches. Train, Validation, and Test datasets were loaded using torch. utils. data. DataLoader(). A pre-trained SWIN-Transformer with unfrozen feature extraction weights was used for training.

Results

Imaged-Based Deep Learning for Healthy vs CLL Classification Outperforms Morphometric Analysis

[0151] Primary human T cells from 3 healthy (H) donors and 7 CLL patients were allowed to spread on hard polydimethylsiloxane (PDMS) surfaces coated with first layer goat-anti-mouse antibody and second layer O-CD3/CD28. These hard substrates were prepared using Sylgard 184 (Dow Silicones Corporation, Midland, Ml, USA) with crosslinker: base = 1 :10. Elastic (Young’s) modulus was measured as 2300 kPa PDMS by Instron device.

[0152] T-cell spreading images on hard PDMS substrates were acquired from 3 healthy donors and 7 CLL patients. Morphological features, including Area, Perimeter, Width, Height, Major, Minor, Circularity, Feret’s Diameter, Aspect Ratio (AR), Roundness, and Solidity, were measured for well-segmented single cells, which were screened by an automated machine learning model. After aggregating the data into image level and balancing the image numbers between categories, there are 167 images in Healthy and 155 images in CLL. Classification between Healthy and CLL was conducted by the established feature-based machine learning approaches mentioned in Example 1. Decision Tree with a single feature (Area) as input resulted in an accuracy of 0.538, an Area Under the Curve (AUC) of 0.540, a sensitivity of 0.559, and a specificity of 0.516. Decision Tree with 11 morphological features as input achieved an accuracy of 0.643, an AUC of 0.641 , a sensitivity of 0.681 , and a specificity of 0.601. Random Forest with all 11 morphological features as input demonstrated an improved accuracy of 0.693, an AUC of 0.768, a sensitivity of 0.724, and a specificity of 0.659 (T able 2.1).

Table 2.1 : Model performance comparison for H vs CLL classification on hard PDMS surface

[0153] To further improve the classification performance, image-based deep learning approach was applied (Figure 17). In preprocessing, each raw image (1002x1004) was broken into 25 (5 rows x 5 columns) smaller image patches (224x224), which resulted in 3887 image patches of CLL patients and 4250 image patches of healthy patients after balancing. The patches were then split into Train, Validation, and Test datasets at a ratio of 80%:10%:10%. In Train dataset, there are 3400 image patches of Healthy and 3109 image patches of CLL; in Validation dataset, there are 425 image patches of Healthy and 288 image patches of CLL; in Test dataset, there are 425 image patches of Healthy and 390 image patches of CLL. Image patches were input into a feature extractor followed by a classifier to classify image patches into Healthy/CLL. The image-level prediction was decided after a majority voting mechanism, which means if more than 50% of the image patches derived from the same image were predicted as one label, then this raw image was predicted as that label.

[0154] The pretrained Swin Transformer model was loaded and trained for 100 epochs, without freezing the feature extraction weights. A learning rate of 1e^-6 was used. The loss and accuracy over 100 epochs were plotted for the train and validation dataset (Figure 18A). The best epoch was selected based on the highest validation accuracy, which was epoch 84, and then was loaded to the test dataset. The model performance on test dataset reached an image-level test accuracy of 0.859, a test AUC of 0.887, a test sensitivity of 0.810, and a test specificity of 0.929 (Table 2.1). The Receiver Operating Characteristic (ROC) curve (Figure 18B) and confusion matrix (Figure 18C) were generated for the test dataset results.

Transfer learning with pretrained Swin Transformer model boosts the model performance compared to unpretrained model

[0155] To evaluate whether a pretrained model is necessary, an unpretrained Swin Transformer model was loaded with the other parameters the same as the pretrained model previously described. The un-pretrained model was trained for 100 epochs, with a learning rate of 1e^-6 (Figure 19A). The model reached an image-level test accuracy of 0.754, a test AUC of 0.783, a test sensitivity of 0.721 , and a test specificity of 0.801 (Table 2.1 ). The Receiver Operating Characteristic (ROC) curve and confusion matrix were generated for the test dataset (Figures 19B-19C). This result validated the importance of transfer learning, providing a significant boost in performance with the pretrained model.

Unfreezing feature extraction weights of the pretrained Swin Transformer model leads to a noticeable improvement compared to freezing the weights [0156] To evaluate the impact of freezing versus unfreezing the feature extraction weights on model performance, a comparative analysis was conducted. In one experiment, the feature extraction weights were frozen by setting “param. requires_grad = False”, effectively preventing any updates during training. The model was then trained for 100 epochs under this condition (Figures 20A-20C). The results were as follows: an accuracy of 0.720, an AUC of 0.887, a sensitivity of 0.642, and a specificity of 0.833 (Table 2.1 ). These metrics suggest that while the model performed reasonably well with frozen feature extraction weights, there is room for improvement, particularly in sensitivity, which indicates the model's ability to correctly identify positive cases.

[0157] In contrast, unfreezing the feature extraction weights, which allows the model to fine-tune these weights to our specific dataset, led to a noticeable improvement in performance. The model became better tailored to the unique characteristics of our data, resulting in enhanced accuracy and overall classification performance.

[0158] The improvement highlights the importance of allowing the model to adapt its feature extraction layers during training, particularly in complex tasks where the dataset may differ significantly from those the model was originally pretrained on. This fine-tuning process enables the model to extract more relevant features, thereby improving its ability to distinguish between classes and increasing both the sensitivity and specificity of its predictions.

Swin Transformer model demonstrates superior capability in extracting relevant features than ResNet-50

[0159] The performance of the Swin Transformer model was compared with the ResNet-50 model on the same task by using both models for feature extraction, keeping the rest of the model architecture consistent such as the classification head. For both models, the feature extraction layers were unfrozen, allowing their weights to be updated during training. This step ensures that the model can fine-tune its feature extraction process according to the specific dataset used in our task, rather than relying solely on the pre-trained weights.

[0160] After training the ResNet-50 model for 100 epochs (Figures 21A-21C), it reached an accuracy of 0.817, AUC of 0.845, sensitivity of 0.801 , and specificity of 0.840 (Table 2.1). These metrics suggest that ResNet-50 performed reasonably well, with a balanced ability to correctly identify positive cases (sensitivity) and negative cases (specificity). However, the performance is still not as good as Swin Transformer. The Swin Transformer's performance might be attributed to its ability to capture long-range dependencies and more complex patterns in the data through its hierarchical structure and attention mechanisms, which are particularly powerful in tasks involving diverse and detailed features. This underscores the importance of model selection based on the nature of the task and the type of data being analyzed.

Imaged-Based Deep Learning for Stiffness Classification Outperforms Morphometric Analysis

[0161] Following the successful classification of T cell intrinsic states (such as distinguishing between healthy and CLL-like cells) using both feature-based machine learning and image-based deep learning, it was further explored whether T cell morphology could also be used to identify extrinsic environmental factors, specifically substrate stiffness.

[0162] A decision tree model was trained using a single morphological feature, Area, to classify the stiffness of the substrate T were spread on, either hard or soft. The model achieved an accuracy of 0.537, AUC of 0.540, sensitivity of 0.511 , and specificity of 0.562. These results indicate that using only one feature provides a classification performance slightly better than random guessing, suggesting that Area alone is not sufficiently informative for this task. The model was then expanded to include multiple morphological features. The multi-feature decision tree improved the classification performance, achieving an accuracy of 0.631 , AUC of 0.630, sensitivity of 0.543, and specificity of 0.716. This improvement highlights the benefit of incorporating additional features, which together provide a more comprehensive representation of T cell morphology under different substrate stiffness conditions. To further enhance the model's predictive power, a random forest classifier was employed, which combines 100 decision trees to improve robustness and accuracy. This model reached an accuracy of 0.693, AUC of 0.755, sensitivity of 0.626, and specificity of 0.758. The random forest's ability to capture complex, non-linear relationships among multiple features contributed to its superior performance compared to the single and multi-feature decision trees. Finally, a pretrained Swin Transformer model was applied to the same classification task. The Swin Transformer significantly outperformed the feature-based models, achieving an accuracy of 0.741 , AUC of 0.761 , sensitivity of 0.750, and specificity of 0.732. The Swin Transformer's ability to analyze the entire cell morphology holistically, rather than relying on pre-extracted features, likely accounts for this enhanced performance.

[0163] These results collectively demonstrated that T cell morphology is not only a useful marker for identifying intrinsic states, such as distinguishing between healthy and CLL-like cells, but can also effectively classify extrinsic environmental factors like substrate stiffness. The transition from simple feature-based models to more sophisticated deep learning approaches underscores the potential of imagebased models to capture subtle morphological differences that are critical for accurate classification. This finding opened up new avenues for using T cell morphology as a versatile tool in both disease diagnosis and in understanding cellular responses to different environmental conditions.

Table 2.2: Model performance comparison for Hard vs Soft classification surface

Discussion

[0164] This Example demonstrated the potential of T cell morphology as a powerful tool for distinguishing both intrinsic cellular states and extrinsic environmental factors. By systematically exploring various machine learning models - from simple decision trees to advanced image-based deep learning techniques - the strengths and limitations of each approach were highlighted, ultimately showing the significant advantages of using deep learning for this type of classification task.

[0165] Building on the morphometric analysis results of Example 1 , this example first tested this analysis with new PDMS formulation, new coating methods and new donors. The classification of Healthy vs CLL on hard surface showed a similar performance trend with Example 1 in that random forest with 11 morphological features as input performed the best among the three models, reaching a moderate classification accuracy. However, one difference in the results between these two Examples was that the predictive power of Area only in classification of disease states is higher in Example 1 than in Example 2, which might result from the variation between two cohorts of CLL patients. This also indicates that a more robust model is needed rather than only relying on single feature such as Area to assess T cell functions for some donors.

[0166] Turning to the impact of extrinsic environmental stiffness on T cell spreading, Example 2 is consistent with Example 1 in that T cells of healthy donors show larger area on softer material (Figures 22A-22B). If combining the two stiffness ranges together, it looks like a biphasic mechanosensing profile where 250-550 kPa showed the peak spreading area. The difference of absolute area in the two Examples might result from the different coating methods. However, for two groups of CLL patients, the mechaonsensing response was different, which makes sense regarding the variability of responsiveness among individuals. Additionally, both Examples show consistent results that mechanosensing is more pronounced in healthy donors.

[0167] To improve and automate the assay, this Example developed an image-based deep learning approach, which achieved a significant leap in performance, especially the Swin Transformer model for feature extraction. The Swin Transformer model's superior accuracy, sensitivity, and specificity underscore the power of deep learning in capturing the intricate details of cell morphology that may be missed by traditional feature analysis methods. By processing the entire image, the Swin Transformer can identify subtle patterns and relationships within the data, leading to more accurate and reliable classifications. [0168] The success of the Swin Transformer in this study suggests that deep learning models could revolutionize the way cellular morphology is analyzed, offering a scalable and highly automated approach that can be applied to large datasets with minimal human intervention. This could be particularly beneficial in clinical settings, where rapid and accurate assessments of cell states and environments are crucial for diagnosis and treatment decisions. Building on these results, it is also aimed to leverage the pretrained model to predict T cell proliferation and improve T cell production by tailoring expansion conditions, as discussed in the next Example.

Example 3 Predicting T Cell Proliferation by Image-Based Deep Learning

[0169] T cells from chronic cancer patients often exhibit deficiencies in proliferation due to T cell exhaustion, posing significant challenges for pretreatment expansion in T cell therapy. To address this issue, recent studies have explored the impact of mechanical properties on ex vivo T cell expansion. It has been reported that replacing stiff materials with softer ones can enhance the expansion of CLL patients’ T cells, although the effects vary among different patients. This variability motivates us to apply our deep learning model to predict T cell proliferation and determine the optimal growth condition for each individual.

[0170] In the deep learning-based regression model used in this disclosure, the input is T cell images collected from our established spreading assay, while the output is max doubling index, a measurement of T cell proliferative capacity. To obtain the ground truth, T cell expansion was conducted under three different stiffness conditions for each donor, including three healthy donors and seven CLL donors. This Example demonstrated a proof-of-principle that T cell spreading images can predict proliferative capacity. This model can be further optimized and generalized to improve its predictive accuracy and applicability across diverse patient populations.

Materials and Methods

PDMS Substrate Preparation forT Cell Proliferation

[0171] PDMS substrates of varying stiffness were prepared by tuning the crosslinker to base ratio. Mass ratios of 1 :10, 1 :30, and 1 :50 (crosslinker: base) produced substrates with Young’s modulus of 2300 kPa (hard), 550 kPa (medium), and 50 kPa (soft). PDMS mixture was poured into a 24-well plate and cured at 65 °C for 16 hours ready for Healthy T cell proliferation assay. For CLL T cell proliferation assay, PDMS was poured into a 48-well plate to downscale the cell number needed due to the limited number of isolated T cells from CLL patients. The same coating method was used for proliferation assay with spreading assay, only to adjust the volume to cover 24-well or 48-well plates.

[0172] Cured PDMS substrates were coated with activating antibodies, first layer goat-anti-mouse antibody and second layer OKT3/9.3, same with the coating method in Example 2. To ensure consistent OKT3/9.3 presentation across different PDMS stiffness, the first layer goat-anti-mouse antibody was adjusted across stiffness, specifically 3ug/ml for hard and medium surfaces and 1 ug/ml for soft surfaces. ELISA was performed to validate the second layer antibody.

Assays of Cell Proliferation

[0173] Healthy T cells were thawed and rested overnight in 37 °C, and CLL T cells were isolated from PBMC and rested overnight in 37 °C. On Day 0, T cells were diluted to 1 x io⁶ cells/mL of complete media, and 1 mL healthy T cell solution was seeded onto the PDMS substrate in a 24 well plate well at a density of 5660 cells/mm². For positive controls, 1 x 10⁶ healthy cells were mixed with 1 x 10⁶ Dynabeads Human T activator CD3/CD28 (Thermo) (25pl) and seeded onto a sterile 24-well plate well.

[0174] On Day 3, cells were removed from the PDMS substrates with gentle pipetting and from Dynabeads with a magnetic holder provided by the vendor. Cell solution was counted and diluted with complete media to a density of 0.5 x 10⁶ cells/mL, and 1 mL of cell solution was reseeded onto new uncoated tissue culture wells. If the volume of fresh media needs to add is lower than 500ul, 500 pd was added to ensure sufficient nutrient. Enumeration and dilution were repeated every 48 hours thereafter until reaching the maximum doubling, a measure representing the proliferative capacity.

[0175] CLL T cell proliferation assay follows the same protocol, only downscaling the volume into 48-well plates. Specifically, 0.5 mL CLL T cell solution was seeded into PDMS substrate in a 48 well plate well, with concentration of 1 M/ml on day 0 and 0.5M/ml from day 3.

Image-Based Deep Learning Model to Predict Proliferation

[0176] The same image dataset with Example 2 was used as input to a regression model to predict maximum doublings. There are 30 true label folders reflecting the proliferation max doubling of 10 donors on 3 different stiffness (Table 3.1 ). A pre-trained SWIN-Transformer was loaded for feature extraction, and the head was modified appropriately for our regression purposes, removing “sigmoidO” from the classification head introduced in Example 2. This model was trained with image patches as input, for 250 epochs using a mean sguared error (MSE) loss function. The best epoch was selected based on the highest accuracy in validation and loaded to test dataset, and R² score was calculated to measure the model performance on test dataset.

[0177] For the test dataset, the scatter plot was generated to show the fit between prediction and true label. Kernel Density Estimation (KDE) was used to highlight the point with the highest density in the predicted data. KDE is a technique used to estimate the probability density function of a continuous random variable. The density of each point was calculated using the KDE model and the index of the point with the highest density was found.

Table 3.1 : Sample list and image patch numbers of 3 Healthy donors and 7 CLL patients

Post-Processing to Output Predicted Optimal Stiffness

[0178] To visualize the predicted max doubling on three different stiffnesses, an overlayed histogram was generated for each donor. The histogram shows the distribution of the predicted max doubling of all the patches. Three stiffness were color-coded: green color for Hard, red color for Medium, and blue color for Soft. Among the three distributions, the most right-forward one means the best stiffness for proliferation.

[0179] To quantify the model performance of predicting the optimal stiffness, the most frequently predicted max doubling value was found as representative of all patches in one stiffness, and three stiffnesses were ranked based on this predicted max doubling value. Accuracy, AUC, sensitivity and specificity of predicting the best stiffness were calculated to measure the performance.

Results

T cells respond to stiffness in both spreading and proliferation

[0180] Primary human T cells from 3 healthy (H) donors and 7 CLL patients were allowed to spread on polydimethylsiloxane (PDMS) surfaces coated with first layer goat-anti-mouse antibody and second layer a-CD3/CD28 (Figure 23A). These substrates were prepared using Sylgard 184 (Dow Silicones Corporation, Midland, Ml, USA). Elastic (Young’s) modulus was modulated by changing the ratio of crosslinker and base, producing three different stiffnesses of 50 (Soft, 1 :50 ratio of crosslinker: base), 1000 (Medium, 1 :30), and 2300 (Hard, 1 :50) kPa PDMS (Figure 23B). Comparison of the absorption from ELISA showed that the concentration of second-layer antibodies was similar across the three different formulations with the adjusted first-layer goat-anti-mouse antibodies, 3| g/ml for Hard and Medium surfaces and 1 g/ml for Soft surface (Figure 23C).

[0181] T cell spreading and proliferation assays were performed on 3 healthy donors and 7 CLL patients. For spreading assay, T cell images were acquired for morphometric analysis and as input for the deep learning model. For proliferation, cells were counted every 2 days until cells stopped growing, usually taking 13-15 days. The results show that both T cell spreading and proliferation respond to stiffness (Figure 24). This indicates that T cell morphology has the potential to predict the optimal stiffness for growth. However, a single feature such as Area is not sufficient to predict proliferation, reflected by the different trends between Area and Max Doubling as a function of stiffness. Image-based deep learning predicts proliferation and optimal growth condition

[0182] T cell images from 3 healthy donors and 7 CLL patients were input to a regression model to predict max doubling on three stiffness conditions. Similar with the classification workflow mentioned in the Example 2, each raw image (1002x1004) was sliced into 25 smaller image patches (224x224), which was input to a pretrained Swin Transformer model for feature extraction. The model head was modified to output a numerical value. The loss between the output and true label was calculated by MSE() loss function (Figure 25). The model was trained for 250 epochs, with the best performance observed at epoch 118, which was loaded for testing on the test dataset, which reached r² score of 0.311 (Figure 26).

[0183] A scatter plot was created to show the relationship between prediction and true labels, with the point with the highest density in the predicted values highlighted by using Kernel Density Estimation (KDE) (explained above). The plot demonstrated a monotonic correlation between predictions and true labels, validating the model’s capability to predict T cell proliferation based on images (Figure 27).

[0184] To predict the optimal stiffness for expansion, an overlayed histogram showing patch predictions of three stiffness conditions was generated (Figure 28) for each donor. The rightmost population in each histogram corresponds to the optimal stiffness condition. To quantitatively rank the three stiffnesses, the predicted value with the peak density for each stiffness was identified and compared. The stiffness condition with the highest predicted value was deemed the best. The model’s prediction reached an accuracy of 0.8, AUC of 0.6875, sensitivity of 0.8, and specificity of 0.9 (Figure 29).

Discussion

[0185] This Example demonstrated that T cell short-term spreading images can serve as an early marker of long-term proliferation. The deep learning-based regression model was used to predict proliferation index with T cell images as input. Although the model’s predictions are monotonically correlated with the true labels, improvement can be made for the model, especially in predicting those with higher proliferation index. More donors can be included to generalize the model. To make the model more robust, instead of directly splitting all the images into train/validation/test datasets, the datasets can be split at the donor level to avoid data leakage. In addition, different stiffness ranges (especially the softer range), substrate formats, and coating methods can be investigated for the broader application of this workflow. An attention map can also be used to visualize where in the image the model is extracting features from so the model can be better interpreted.

[0186] This Example provided a strong proof of concept, and further improvements can be made in the following avenues. First, the dataset can be expanded to include a broader range of environmental conditions and cell types, which would help validate the generalizability. Additionally, other advanced deep learning models or hybrid approaches that combine deep learning with traditional feature-based methods could be employed to further enhance the accuracy and interpretability of the results. Moreover, integrating images with other data types, such as gene expression profiles or proteomics data, could provide a more comprehensive understanding of the underlying biological processes. Finally, the development of real-time analysis tools that can monitor T cell behavior dynamically, in addition to static images, could open up new possibilities for studying cellenvironment interactions in more physiologically relevant contexts. For example, time-lapse imaging combined with deep learning could be used to track the progression of cellular responses to environmental changes, providing insights into the temporal dynamics of cell behavior.

Example 4

Investigating Short-Term Mechanosensing Response in T Cell Subsets

[0187] T cells play a critical role in the immune response and are composed of various subsets, each with unique functions. The composition of these subsets can vary significantly across donors due to factors such as age, sex, genetic background, and health status. Understanding the variability in T cell subset composition is essential for optimizing immunotherapies, including cell-based therapies and personalized treatments.

[0188] Major T cell subsets include naive cells, central memory cells, effector memory cells, and effector cells. Naive cells have not yet encountered specific antigens. They circulate through the peripheral lymphoid organs, such as the lymph nodes and spleen, where they constantly survey for the presence of new infections. Upon encountering their specific antigen presented by antigen-presenting cells (APCs), naive T cells become activated, initiating a primary immune response. After a naive T cell encounters its antigen and becomes activated, it can differentiate into a central memory T cell. These cells have encountered antigens previously but remain in a less differentiated state compared to effector cells. Central memory T cells reside primarily in secondary lymphoid organs and are characterized by their high proliferative capacity and ability to rapidly differentiate into effector cells upon re-exposure to the same antigen. This enables a swift and robust response during subsequent infections. Effector memory T cells also have encountered antigens, but unlike central memory T cells, they do not express lymphoid homing receptors (e.g., CCR7). Instead, they circulate in peripheral tissues, such as the skin and mucosa, where they can quickly respond to infections at the site of pathogen entry. Effector memory T cells are more differentiated than central memory T cells and are poised to exert effector functions, such as producing cytokines or exerting cytotoxicity activity upon reactivation. Effector T cells are fully differentiated cells that arise during an active immune response. Effector T cells can be either cytotoxic T cells, which directly kill infected or tumor cells, or helper T cells, which coordinate the immune response by activating other immune cells. Effector T cells are crucial for the immediate defense against infections and tumors. Once their job is done, most effector T cells undergo apoptosis, but some persist as memory T cells to provide long-term immunity.

[0189] Previous studies have demonstrated that T cells from different donors exhibit variability in their response to mechanical properties. To optimize T cell expansion for cell therapy, the aim of this Example is to investigate the role of T cell subtypes in mechanosensing. Building on the established T cell spreading assay, subset markers were stained to analyze cell morphology of each T cell subset.

Materials and Methods

T Cell Spreading and Immunostaininq

[0190] T cells were seeded onto glass-supported PDMS substrates at a concentration of 1 x 10⁶ cells/mL. Following 40 min T cell spreading, samples were fixed and permeabilized using True-Nuclear Transcription Factor Buffer Set. Alexa Fluor (AF) 488 phalloidin (1 :40 dilution), PerCP anti-human CD4 (clone: RPA-T4, 1 :40 dilution), AF 647 anti-human CD45RA (clone: H1100, 1 :50 dilution), and PE antihuman CCR7 (clone: FR 1 1-1 1 E8, 1 :50 dilution) were added in perm buffer to stain the cells for 30 min at room temperature (Figure 30). After staining, cells were washed by PBS and imaged under 40* magnification using an Olympus 1X81 inverted microscope, equipped with an Andor iXon EMCCD camera, providing a 1002 x 1004 array of 8 pm x 8 pm pixels. Cell morphological features were measured under Alexa Fluor (AF) 488 phalloidin channel by Imaged. The segmented cells were then applied to other channels to measure the median intensity. Data analysis was performed on Excel, plotting histogram of the cell median intensity in each channel and filtering positive or negative cells by thresholding.

Flow Cytometry

[0191] Purity of T cells were verified by Live/Dead (L/D) Fixable Violet and FITC anti-human CD3 antibody via FACSCanto I. T cells were also stained with L/D FITC, PerCP anti-human CD4, AF 647 anti-human CD45RA, and PE anti-human CCR7, and were assessed via Cytek Aurora to collect baseline composition of T cell subsets (Figures 31A-31 B).

Results

Different donors show different T cell compositions from flow cytometry characterization

[0192] T cell immunophenotyping was performed on samples from 3 healthy donors and 7 CLL patients (Figures 32A-32D). The result indicated variability in CD47CD8⁺ T cell ratios among different donors, reflecting individual immune system differences. Additionally, the proportion of four T cell subtypes (naive, central memory, effector memory, and effector) varied across donors. Notably, CLL patients exhibited a significantly higher percentage of memory T cells compared to healthy donors, which may be indicative of an immune response to persistent antigen exposure. When comparing T cell subsets within the CD4⁺ and CD8⁺ populations, both healthy donors and CLL patients showed a higher percentage of effector cells and a lower percentage of central memory T cells within the CD8⁺ population, and CLL patients exhibited more effector cells than healthy donors. This distribution suggests a shift towards a more differentiated and potentially cytotoxic T cell profile in the context of CLL.

Visualization of T cell morphology and subset identification from fixed imaging

[0193] T cells were allowed to spread on PDMS with varied stiffness for 40 min, followed by fixation, permeabilization, and staining with antibodies to actin, CD4, CD45RA, and CCR7. Images of four fluorescence channels were acguired under 40x magnification (Figures 33A-33B). To characterize cell subsets, naive T cells are CD45RA7CCR7⁺, central memory T cells are CD45RA7CCR7⁺, effector memory T cells are CD45RA7CCR7', effector T cells are CD45RA7CCR7'.

Mechanosensitivity of CD4⁺ and CD8⁺ T cells varies among different donors

[0194] Mechanosensitivity of CD4⁺ and CD8⁺ T cells was assessed by measuring the spreading area across different substrate stiffness for multiple donors. CD4⁺ and CD8⁺ T cells were distinguished using PerCP anti-human CD4 staining. The results revealed significant inter-donor variability in mechanosesnsing responses. For Healthy Donor 1 , 2 and CLL1 , CD4⁺ T cells exhibited greater sensitivity to substrate stiffness, showing a smaller p-value. Conversely, in Healthy Donor 3 and CLL3, CD8⁺ T cells demonstrated higher mechanosensitivity. This variation underscores the individualized nature of T cell response to mechanical cues, likely influenced by inherent genetic and environmental factors. Additionally, the decreased CD4⁺ occupation in Healthy Donor 3 and CLL 3 KeOb (as shown in Figure 34) may result in a deficiency in CD4⁺ T cell mechanosensing. It could be that the dominance of CD4⁺ T cells in the population leads to its higher capability to sense the stiffness. Only when the percentage of CD4⁺ T cells significantly drops, CD8⁺ T cells would show higher sensitivity to stiffness.

[0195] These findings highlighted the necessity of considering donor-specific differences in mechanobiology studies and suggest that mechanosensitivity in T cells may not be uniformly distributed across all individuals.

Memory T cells show a larger spreading area and T cell subsets present a more varied mechanosensitivity in CLL patients

[0196] The mechanosensitivity of various T cell subsets - Naive, Central Memory (CM), Effector Memory (EM), and Effector T cells - was evaluated by analyzing their spreading areas on substrates of different stiffnesses in samples from two healthy donors and one CLL patient (Figure 35). The results revealed notable differences in how these T cell subsets respond to mechanical cues, which may provide insights into disease-related alterations in cellular behavior.

[0197] Memory T cells, particularly Effector Memory (EM) T cells, exhibited a significantly larger spreading area compared to other T cell subsets across different substrate stiffnesses, potentially reflecting their readiness to respond rapidly to reinfection or re-exposure to antigens by spreading. Interestingly, the spreading area of CM T cells was significantly reduced in the CLL patient compared to the healthy donors. This suggests that CM T cells in CLL may be impaired in their ability to form immune synapse, potentially contributing to the distinct morphological differences observed between healthy and CLL T cells. Since CM T cells play a crucial role in maintaining long-term immunity and responding to re-infection, their altered spreading in CLL could be a key factor in the disease's progression and the patient's immune dysfunction. Moreover, since CM T cells are characterized by their high proliferative capacity, the impaired spreading of CM T cells can be an indicator of the proliferation deficiency in CLL patients, which provides another layer of evidence that T cell morphology has the potential to predict long-term proliferation.

[0198] Another notable observation is the difference in mechanosensing trends among T cell subsets between healthy donors and the CLL patient. In healthy donors, T cell subsets displayed a relatively uniform mechanosensitive response, indicating a consistent adaptation to varying substrate stiffnesses. In contrast, the CLL patient exhibited more variable mechanosensing among T cell subsets, suggesting that the disease may disrupt the typical mechanosensory coordination among T cells. This variability could be a reflection of the altered immune landscape in CLL, where dysregulated T cell function and interaction with the tumor microenvironment may lead to heterogeneous responses to mechanical cues.

Discussion

[0199] This Example investigated the mechanosensing response of T cells by decoupling it across different subsets. The findings indicated that CD4⁺ T cells exhibit a dominant mechanosensitive response compared to CD8⁺ T cells. However, this dominance appears to be influenced by the relatively lower percentage of CD4⁺ T cells within the sample. This observation warrants further investigation, particularly by including a larger cohort of donors to determine if this trend is consistent across diverse populations.

[0200] The comparison of mechanosensitivity among Naive, Central Memory (CM), Effector Memory (EM), and Effector T cells has provided deeper insights into the differences between healthy donors and CLL patients. In healthy donors, T cell subsets exhibited a uniform mechanosensitive response, reflecting coordinated behavior across the different T cell types. In contrast, CLL patients demonstrated a more variable mechanosensory response among T cell subsets. This variance suggests that CLL may disrupt the typical mechanosensory coordination, potentially leading to functional impairments in T cell responses and contributing to disease progression.

[0201] While these findings are intriguing, it is important to acknowledge the limitations of the current study. One significant limitation was the effectiveness of CCR7 staining, which was not consistently reliable in distinguishing between positive and negative populations during imaging. This staining issue could lead to inaccuracies in the identification and classification of the T cell subsets, particularly affecting the interpretation of mechanosensitivity among Naive, CM, EM, and Effector T cells.

[0202] To address this, further optimization of the immunostaining protocol is necessary to ensure clear and accurate separation of the four T cell subsets in imaging studies. Such improvements are critical for validating our findings and ensuring that the observed differences in mechanosensitivity are indeed reflective of intrinsic properties of the T cell subsets and not artifacts of suboptimal staining. Future studies should also consider expanding the number of donors and patient samples to validate these mechanosensing trends in a broader context. Additionally, these channels of images to identify T cell subsets can be added as input for optimizing our deep learning model to predict long-term proliferation and other functional outcomes. CITED DOCUMENTS

1. “Definition of CAR T-cell therapy - NCI Dictionary of Cancer Terms - NCI.”

Accessed: Aug. 03, 2024. [Online], Available: https://www.cancer.gov/publications/dictionaries/cancer-terms/def/car-t-cell- therapy

2. A. Wahl et al., “Biphasic mechanosensitivity of T cell receptor-mediated spreading of lymphocytes,” Proceedings of the National Academy of Sciences, vol. 116, no. 13, pp. 5908-5913, Mar. 2019, doi: 10.1073/pnas.1811516116.

3. R. S. O’Connor et al., “Substrate rigidity regulates human T cell activation and proliferation,” J Immunol, vol. 189, no. 3, pp. 1330-1339, Aug. 2012, doi: 10.4049/jimmunol.1102757.

4. W. Koehrsen, “Random Forest Simple Explanation,” Medium. Accessed: Aug. 06, 2024. [Online], Available: https://williamkoehrsen.medium.com/random- forest-simple-explanation-377895a60d2d

5. A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” Commun. ACM, vol. 60, no. 6, pp. 84- 90, May 2017, doi: 10.1145/3065386.

6. Z. Liu et al., “Swin Transformer: Hierarchical Vision Transformer using Shifted

Windows,” Aug. 17, 2021 , arXiv arXiv:2103.14030. doi:

10.48550/arXiv.2103.14030.

7. J. H. Lee, S. Shao, M. Kim, S. M. Fernandes, J. R. Brown, and L. C. Kam, “Multi-Factor Clustering Incorporating Cell Motility Predicts T Cell Expansion Potential,” Front Cell Dev Biol, vol. 9, p. 648925, 2021 , doi: 10.3389/fcell.2021.648925.

8. R. N. Palchesko, L. Zhang, Y. Sun, and A. W. Feinberg, “Development of polydimethylsiloxane substrates with tunable elastic modulus to study cell mechanobiology in muscle and nerve,” PLoS One, vol. 7, no. 12, p. e51499, 2012, doi: 10.1371/journal. pone.0051499. P. K. Chaudhuri, M. S. Wang, C. T. Black, M. Huse, and L. C. Kam, “Modulating T Cell Activation Using Depth Sensing Topographic Cues,” Adv Biosyst, vol. 4, no. 9, p. e2000143, Sep. 2020, doi: 10.1002/adbi.202000143. A. Boussommier-Calleja, R. Li, M. B. Chen, S. C. Wong, and R. D. Kamm, “Microfluidics: A new tool for modeling cancer-immune interactions,” Trends Cancer, vol. 2, no. 1 , pp. 6-19, Jan. 2016, doi: 10.1016/j.trecan.2015.12.003. K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” Apr. 10, 2015, arXiv arXiv:1409.1556. doi: 10.48550/arXiv.1409.1556. C. Szegedy et al., “Going Deeper with Convolutions,” Sep. 16, 2014, arXiv. arXiv: 1409.4842. doi: 10.48550/arXiv.1409.4842. K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2016, pp. 770-778. doi: 10.1109/CVPR.2016.90. “What is ResNet-50?,” Roboflow Blog. Accessed: Sep. 03, 2024. [Online], Available: https://blog.roboflow.com/what-is-resnet-50/ D. L. Porter, B. L. Levine, M. Kalos, A. Bagg, and C. H. June, “Chimeric antigen receptor-modified T cells in chronic lymphoid leukemia,” N Engl J Med, vol. 365, no. 8, pp. 725-733, Aug. 2011 , doi: 10.1056/NEJMoal 103849. A. G. Ramsay et al., “Chronic lymphocytic leukemia T cells show impaired immunological synapse formation that can be reversed with an immunomodulating drug,” J Clin Invest, vol. 118, no. 7, pp. 2427-2437, Jul. 2008, doi: 10.1172/JCI35017. J. C. Riches et al., “T cells from CLL patients exhibit features of T-cell exhaustion but retain capacity for cytokine production,” Blood, vol. 121 , no. 9, pp. 1612-1621 , Feb. 2013, doi: 10.1182/blood-2012-09-457531 . M. Palma et al., “T cells in chronic lymphocytic leukemia display dysregulated expression of immune checkpoints and activation markers,” Haematologica, vol. 102, no. 3, pp. 562-572, Mar. 2017, doi: 10.3324/haematol.2016.151100. S. Scrivener, E. R. Kaminski, A. Demaine, and A. G. Prentice, “Analysis of the expression of critical activation/interaction markers on peripheral blood T cells in B-cell chronic lymphocytic leukaemia: evidence of immune dysregulation,” Br J Haematol, vol. 112, no. 4, pp. 959-964, Mar. 2001 , doi: 10.1046/j.1365- 2141.2001.02672.x. J. A. Burger and J. G. Gribben, “The microenvironment in chronic lymphocytic leukemia (CLL) and other B cell malignancies: insight into disease biology and new targeted therapies,’’ Semin Cancer Biol, vol. 24, pp. 71-81 , Feb. 2014, doi: 10.1016/j.semcancer.2013.08.011 . F. J. Gassner et al., “Chronic lymphocytic leukaemia induces an exhausted T cell phenotype in the TCL1 transgenic mouse model,” Br J Haematol, vol. 170, no. 4, pp. 515-522, Aug. 2015, doi: 10.1111/bjh.13467. J. Wu et al., “Phenotypic alteration of CD8+ T cells in chronic lymphocytic leukemia is associated with epigenetic reprogramming,” Oncotarget, vol. 7, no. 26, pp. 40558-40570, Jun. 2016, doi: 10.18632/oncotarget.9941. W. Lin et al., “Morphological change of CD4(+) T cell during contact with DC modulates T-cell activation by accumulation of F-actin in the immunology synapse,” BMC Immunol, vol. 16, p. 49, Aug. 2015, doi: 10.1186/s12865-015- 0108-x. A. Grakoui et al., “The immunological synapse: a molecular machine controlling T cell activation,” Science, vol. 285, no. 5425, pp. 221-227, Jul. 1999, doi: 10.1126/science.285.5425.221 . E. Cretel, D. Touchard, P. Bongrand, and A. Pierres, “A new method for rapid detection of T lymphocyte decision to proliferate after encountering activating surfaces,” J Immunol Methods, vol. 364, no. 1-2, pp. 33-39, Feb. 2011 , doi: 10.1016/j.jim.2O10.10.007. E. Judokusumo, E. Tabdanov, S. Kumari, M. L. Dustin, and L. C. Kam, “Mechanosensing in T Lymphocyte Activation,” Biophys J, vol. 102, no. 2, pp. L5-L7, Jan. 2012, doi: 10.1016/j.bpj.2O11 .12.011. M. Saitakis et al., “Different TCR-induced T lymphocyte responses are potentiated by stiffness with variable sensitivity,” Elife, vol. 6, p. e23190, Jun. 2017, doi: 10.7554/eLife.23190. C. Sachar and L. C. Kam, “Probing T Cell 3D Mechanosensing With Magnetically-Actuated Structures,” Front Immunol, vol. 12, p. 704693, 2021 , doi: 10.3389/fimmu.2021.704693. D. J. Yuan, L. Shi, and L. C. Kam, “Biphasic response of T cell activation to substrate stiffness,” Biomaterials, vol. 273, p. 120797, Jun. 2021 , doi: 10.1016/j.biomaterials.2021.120797. “Citing,” ImageJ Wiki. Accessed: Jan. 28, 2024. [Online], Available: https://imagei.github.io/contribute/citinq J. R. Quinlan, “Induction of decision trees,” Mach Learn, vol. 1 , no. 1 , pp. 81 — 106, Mar. 1986, doi: 10.1007/BF00116251 . L. Breiman, “Random Forests,” Machine Learning, vol. 45, no. 1 , pp. 5-32, Oct. 2001 , doi: 10.1023/A: 1010933404324. L. Breiman, Classification and Regression Trees. New York: Routledge, 2017. doi: 10.1201/9781315139470. A. Dang et al., “Enhanced activation and expansion of T cells using mechanically soft elastomer fibers,” Adv Biosyst, vol. 2, no. 2, p. 1700167, Feb. 2018, doi: 10.1002/adbi.201700167. X. Wang, S. M. Fernandes, J. R. Brown, and L. C. Kam, “Assaying and Classifying T Cell Function by Cell Morphology,” BioMedinformatics, vol. 4, no. 2, Art. no. 2, Jun. 2024, doi: 10.3390/biomedinformatics4020063. N. Leibovitch, “Exploring Visual Attention in Transformer Models,” Medium.

Accessed: Oct. 03, 2024. [Online], Available: https://medium.com/@nivonl/exploring-visual-attention-in-transformer-models- ab538c06083a R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization,” Int J Comput Vis, vol. 128, no. 2, pp. 336-359, Feb. 2020, doi: 10.1007/sl 1263-019-01228-7.

38. V. Golubovskaya and L. Wu, “Different Subsets of T Cells, Memory, Effector Functions, and CAR-T Immunotherapy,” Cancers (Basel), vol. 8, no. 3, p. 36, Mar. 2016, doi: 10.3390/cancers8030036.

39. “Central Memory and Effector Memory T Cell Subsets: Function, Generation, and Maintenance | Annual Reviews.” Accessed: Sep. 05, 2024. [Online], Available: https://www.annualreviews.orq/content/ioumals/10.1146/annurev.immunol.22. 012703.104702.

40. “U.S. FDA Approves Bristol Myers Squibb’s Breyanzi ® as the First and Only

CAR T Cell Therapy for Adults with Relapsed or Refractory Chronic Lymphocytic Leukemia (CLL) or Small Lymphocytic Lymphoma (SLL).” Accessed: Oct. 03, 2024. [Online], Available: https://news.bms.com/news/details/2024/U.S.-FDA-Approves-Bristol-Myers- Squibbs-Breyanzi-as-the-First-and-Only-CAR-T-Cell-Therapy-for-Adults-with- Relapsed-or-Refractory-Chronic-Lymphocytic-Leukemia-CLL-or-Small- Lymphocytic-Lymphoma-SLL/default.aspx

[0203] All documents cited in this application are hereby incorporated by reference as if recited in full herein.

[0204] The embodiments described in this disclosure can be combined in various ways. Any aspect or feature that is described for one embodiment can be incorporated into any other embodiment mentioned in this disclosure. While various novel features of the inventive principles have been shown, described and pointed out as applied to particular embodiments thereof, it should be understood that various omissions and substitutions and changes may be made by those skilled in the art without departing from the spirit of this disclosure. Those skilled in the art will appreciate that the inventive principles can be practiced in other than the described embodiments, which are presented for purposes of illustration and not limitation.

Claims

WHAT IS CLAIMED IS:

1. A method for rapid assessing the health of a cell from a subject in need thereof, comprising the steps of:

(a) training a model that provides a standard reference indicating the health of the cell, comprising: i. isolating the cell from a healthy subject and expanding them on a functionalized substrate; ii. obtaining raw images for the cell by immunofluorescence and fixed imaging; iii. processing the raw images to generate sliding patches; iv. measuring a number of parameters based on the sliding patches; v. assigning an indicator of health to the cell and recording the measured values for the parameters; vi. repeating steps (a-i) to (a-v) with the cell from a number of healthy subjects; vii. consolidating, analyzing and adjusting the measured values for the parameters from all tested healthy subjects to obtain a classifier for the healthy cell; viii. repeating steps (a-i) to (a-vii) by isolating the cell from a subject with a condition and expanding them on the same functional substrate, wherein the cell from the subject with the condition receives an indicator of the condition and obtains a classifier for the cell having the condition; ix. repeating step (a-viii) by isolating the cell from a subject with a different condition and expanding them on the same functional substrate; and x. obtaining the standard reference by consolidating and analyzing the classifiers for the healthy cell and the cell having any condition,

(b) assessing the health of the cell from a subject in need thereof, comprising: i. isolating the cell from the subject in need thereof and obtaining the measured values for the parameters as described in steps (a-i) to (a-iv); ii. entering the measured values in step (b-i) into the trained model; and iii. receiving the indicator of health, and

(c) selecting a suitable cell expansion plan for the subject.

2. A method for predicting the optimal growth condition for expansion of a cell from a subject in need thereof, comprising the steps of:

(a) training a model that provides a standard reference predicting the optimal growth condition for expansion of the cell, comprising: i. isolating the cell from a healthy subject and expanding them on a functionalized substrate; ii. obtaining raw images for the cell by immunofluorescence and fixed imaging; iii. processing the raw images to generate sliding patches; iv. measuring a number of parameters based on the sliding patches; v. assigning an expansion indicator to the cell and recording the measured values for the parameters; vi. repeating steps (a-i) to (a-v) with the cell from a number of healthy subjects; vii. consolidating, analyzing and adjusting the measured values for the parameters from all tested healthy subjects to obtain an expansion classifier for the healthy cell on the functionalized substrate; viii. repeating steps (a-i) to (a-vii) by expanding the cell from healthy subjects on a different functionalized substrate to receive a different expansion indicator and obtain an expansion classifier on the different functionalized substrate; ix. changing the functionalized substrate and repeating steps (a- viii); x. repeating steps (a-i) to (a-ix) by isolating the cell from a subject with a condition; xi. repeating step (a-x) by isolating the cell from a subject with a different condition; and xii. obtaining the standard reference by consolidating and analyzing the classifiers for the healthy cell and the cell having any condition,

(b) predicting the optimal growth condition for expansion of the cell from a subject in need thereof, comprising: i. isolating the cell from the subject in need thereof and obtaining the measured values for the parameters as described in steps (a-i) to (a-iv); ii. entering the measured values in step (b-i) into the trained model; and iii. receiving the expansion indicator, and

(c) selecting an optimal cell expansion plan for the subject.

3. The method of claim 1 or 2, wherein the cell is a lymphocyte.

4. The method of claim 1 or 2, wherein the cell is selected from a T cell, a B cell, a natural killer (NK) cell, and a natural killer T (NKT) cell.

5. The method of claim 1 or 2, wherein the cell is a T cell.

6. The method of claim 2, wherein the optimal growth condition for the cell expansion includes the stiffness, format, and/or coating of the functionalized substrate.

7. The method of any one of claims 1-6, wherein the parameters are morphology features selected from the group consisting of area, roundness, major, minor, perimeter, solidity, circularity, height, width, Feret’s diameter, aspect ratio, and combinations thereof.

8. The method of claim 7, wherein the parameters further comprise a proliferation index.

9. The method of claim 8, wherein the proliferation index comprises a max doubling value.

10. The method of claim 8, wherein the proliferation index is obtained by training a regression model using one or more parameters selected from the group consisting of area, roundness, major, minor, perimeter, solidity, circularity, height, width, Feret’s diameter, aspect ratio, and combinations thereof.

11. The method of claim 7, wherein the parameters further comprise a cell subtype marker and/or a cytoskeletal component.

12. The method of claim 11 , wherein the cell subtype marker is selected from the group consisting of T cell receptor (TCR), CD3, CD4, CD8, FoxP3, CD28, CD45RA, CD45RO, CD62L, CCR7, CD27, CD28, CD25, CD127, CD57, CD137; CD19, CD24, CD38, CD40, CD1 D, IgM; CD11b, CD27, CD161 , CCR7, CD244, NCR3, CD94, CD122, and combinations thereof.

13. The method of claim 11 , wherein the cytoskeletal component is selected from microtubules, intermediate filaments, microfilaments, and combinations thereof.

14. The method of claim 1 or 2, wherein the raw images in step (a-ii) are further obtained by time-lapse videos.

15. The method of claim 14, wherein the parameters further comprise temporal features selected from speed and direction of cell movement.

16. The method of any one of claims 1-15, wherein the condition is selected from an autoimmune disease, fibrosis, a viral infection, a transplant rejection, and a cancer.

17. The method of claim 16, wherein the condition is acute lymphoblastic leukemia (ALL) or chronic lymphocytic leukemia (CLL).

18. A method for treating or ameliorating the effects of a disease in a subject in need thereof, comprising: (a) assessing the health of T cells isolated from the subject according to the method of claim 1 ;

(b) selecting an optimal growth condition for T cells expansion for the subject according to the method of claim 2;

(c) expanding T cells for the subject under the optimal growth condition of step (b) to a clinically relevant number;

(d) reinfusing the expanded T cells to the subject;

(e) after a course of treatment, reassessing the health of T cells in the subject by repeating step (a);

(f) if the reassessment result in step (e) is acceptable, continuing expanding T cells under the current optimal growth condition, or if the reassessment result in step (e) is unacceptable, adjusting T cells expansion by selecting a different optimal growth condition by repeating step (b) and expanding T cells under the adjusted optimal growth condition;

(g) reinfusing the expanded T cells in step (f) to the subject; and

(h) repeating steps (e) to (g) as necessary.

19. The method of claim 18, wherein the disease is selected from the group consisting of an autoimmune disease, fibrosis, a viral infection, a transplant rejection, a cancer, or combinations thereof.

20. The method of claim 19, wherein the autoimmune disease is selected from the group consisting of type 1 diabetes, systemic lupus erythematosus, Sjogren’s syndrome, diffuse scleroderma, inflammatory myopathy, ANCA- associated systemic vasculitis, antiphospholipid syndrome, mucosal-dominant pemphigus vulgaris, anti-MuSK-antibody-positive myasthenia gravis, generalised myasthenia gravis, lupus nephritis, neuromyelitis optica spectrum disorder, myasthenia gravis, chronic inflammatory demyelinating, polyradiculoneuropathy, immune-mediated necrotising myopathy, immune nephritis, refractory POEMS syndrome, amyloidosis, autoimmune haemolytic anaemia, vasculitis, Crohn’s disease, ulcerative Colitis, dermatomyositis, and Still disease.

21. The method of claim 19, wherein the viral infection is caused by a virus selected from human immunodeficiency virus (HIV), hepatitis B virus (HBV), hepatitis C virus (HCV), and cytomegalovirus (CMV).

22. The method of claim 19, wherein the transplant rejection is selected from HLA-A2 mismatched liver transplantation, and HLA-A2 mismatched living donor kidney transplantation.

23. The method of claim 19, wherein the cancer is a hematologic cancer selected from the group consisting of acute lymphoblastic leukemia (ALL), chronic lymphocytic leukemia (CLL), follicular lymphoma, mantle cell lymphoma, diffuse large B-cell lymphoma, and multiple myeloma.

24. The method of claim 19, wherein the cancer is a solid tumor selected from the group consisting of glioblastoma, ependymoma, medulloblastoma, pediatric brain tumor, breast cancer, neuroblastoma, liver cancer, pancreatic cancer, prostate cancer, lung cancer, gastric cancer, and esophageal cancer.

25. The method of claim 18, wherein the disease is chronic lymphocytic leukemia (CLL).

26. The method of claim 18, wherein the subject is a mammal.

27. The method of claim 18, wherein the subject is a human.

28. The method of claim 18, wherein the cell used for treatment is a lymphocyte other than a T cell.

29. The method of claim 18, wherein the cell used for treatment is a natural killer (NK) cell or a natural killer T (NKT) cell.

30. The method of claim 18, further comprising administering to the subject one or more standard therapy for the disease.