Label-Free Identification of Lymphocyte Subtypes Using Three-Dimensional Quantitative Phase Imaging and Machine Learning

Summary

We describe a protocol for the label-free identification of lymphocyte subtypes using quantitative phase imaging and a machine learning algorithm. Measurements of 3D refractive index tomograms of lymphocytes present 3D morphological and biochemical information for individual cells, which is then analyzed with a machine-learning algorithm for identification of cell types.

Abstract

We describe here a protocol for the label-free identification of lymphocyte subtypes using quantitative phase imaging and machine learning. Identification of lymphocyte subtypes is important for the study of immunology as well as diagnosis and treatment of various diseases. Currently, standard methods for classifying lymphocyte types rely on labeling specific membrane proteins via antigen-antibody reactions. However, these labeling techniques carry the potential risks of altering cellular functions. The protocol described here overcomes these challenges by exploiting intrinsic optical contrasts measured by 3D quantitative phase imaging and a machine learning algorithm. Measurement of 3D refractive index (RI) tomograms of lymphocytes provides quantitative information about 3D morphology and phenotypes of individual cells. The biophysical parameters extracted from the measured 3D RI tomograms are then quantitatively analyzed with a machine learning algorithm, enabling label-free identification of lymphocyte types at a single-cell level. We measure the 3D RI tomograms of B, CD4+ T, and CD8+ T lymphocytes and identified their cell types with over 80% accuracy. In this protocol, we describe the detailed steps for lymphocyte isolation, 3D quantitative phase imaging, and machine learning for identifying lymphocyte types.

Introduction

Lymphocytes can be classified into various subtypes including B, helper (CD4+) T, cytotoxic (CD8+) T, and regulatory T cells. Each lymphocyte type has a different role in the adaptive immune system; for example, B lymphocytes produce antibodies, whereas T lymphocytes detect specific antigens, eliminate abnormal cells, and regulate B lymphocytes. Lymphocyte function and regulation is tightly controlled by and related to various diseases including cancers¹, autoimmune diseases², and viral infections³. Thus, the identification of lymphocyte types is important to understand their pathophysiological roles in such diseases and for immunotherapy in clinics.

Currently, methods for classifying lymphocyte types rely on antigen-antibody reactions by targeting specific surface membrane proteins or surface markers⁴. Targeting surface markers is a precise and accurate method to determine lymphocyte types. However, it requires expensive reagents and time-consuming procedures. Furthermore, it carries risks of the modification of membrane protein structures and the alteration of cellular functions.

To overcome these challenges, the protocol described here introduces the label-free identification of lymphocyte types using 3D quantitative phase imaging (QPI) and machine learning⁵. This method enables the classification of lymphocyte types at a single-cell level based on morphological information extracted from label-free 3D imaging of individual lymphocytes. Unlike conventional fluorescence microscopy techniques, QPI utilizes refractive index (RI) distributions (intrinsic optical properties of live cells and tissues) as optical contrast^6,7. The RI tomograms of individual lymphocytes represent phenotypic information specific to subtypes of lymphocytes. In this case, to systemically utilize 3D RI tomograms of individual lymphocytes, a supervised machine learning algorithm was utilized.

Using various QPI techniques, the 3D RI tomograms of cells have been actively used for the study of cell pathophysiology because they provide a label-free, quantitative imaging capability^{8,9,10,11,12,13}. Also, the 3D RI distributions of individual cells can provide morphological, biochemical, and biomechanical information about cells. 3D RI tomograms have been previously utilized in the fields of hematology^14,15,16,17, infectious diseases^18,19,20, immunology²¹, cell biology^22,23, inflammation²⁴, cancer²⁵, neuroscience^26,27, developmental biology²⁸, toxicology²⁹, and microbiology^12,30,31,32.

Although 3D RI tomograms provide detailed morphological and biochemical information of cells, the classification of lymphocyte subtypes is difficult to achieve by simply imaging 3D RI tomograms⁵. To systematically and quantitatively exploit the measured 3D RI tomograms for the cell type classification, we utilized a machine learning algorithm. Recently, several works have been reported in which quantitative phase images of cells were analyzed with various machine learning algorithms³³, including the detection of microorganisms³⁴, classification of bacterial genus^35,36, rapid and label-free detection of anthrax spores³⁷, automated analysis of sperm cells³⁸, analysis of cancer cells^39,40, and detection of macrophage activation⁴¹.

This protocol provides detailed steps to perform label-free identification of lymphocyte types at the individual cell level using 3D QPI and machine learning. This includes: 1) lymphocyte isolation from mouse blood, 2) lymphocyte sorting via flow cytometry, 3) 3D QPI, 4) quantitative feature extraction from 3D RI tomograms, and 5) supervised learning for identifying lymphocyte types.

Protocol

Animal care and experimental procedures were performed under the approval of the Institutional Animal Care and Use Committee of KAIST (KA2010-21, KA2014-01, and KA2015-03). All the experiments in this study were carried out in accordance with the approved guidelines.

1. Lymphocyte Isolation from Mouse Blood

1. Once a C57BL/6J mouse is euthanized via CO₂ inhalation, insert a 26-G needle into the mouse heart and collect 0.3 mL of blood. Directly put blood into a tube with 100 U/mL heparin solution diluted with 1 mL of phosphate-buffered saline (PBS).
   NOTE: Alternatively, lymphocytes from the spleen can be isolated.
2. Centrifuge the tube at 400 x g for 5 min at 4 °C.
3. Add 0.5 mL of ammonium-chloride-potassium lysing buffer to the tube and gently invert it a few times to mix the solution.
4. Incubate the tube at room temperature (RT) for 5 min.
5. Wash the cells by adding 4.5 mL of PBS and centrifuging at 400 x g for 5 min at 4 °C, twice.
6. Remove the supernatant and resuspend the cell pellet in 100 µL of fresh RPMI-1640 medium with 10% fetal bovine serum (FBS).
7. Add 0.1 µg of CD16/32 (2.4G2) antibody to the tube for blocking.
8. Keep the tube on ice.

2. Flow Cytometry and Sorting of Lymphocyte Subtypes

NOTE: Sorting lymphocytes depending on cell type is essential for establishing the ground-truth (i.e., correct) cell type labels to train and test a cell type classifier in supervised learning. Flow cytometry, a gold standard method, is used to identify and separate lymphocytes⁴².

1. Make a mixture of surface marker staining antibodies in 100 µL of fresh RPMI-1640 medium [with 10% FBS, 0.1 µg of CD3e (17A2), CD8a (53-6.7), CD19 (1D3), CD45R (B220, RA3-6B2), and NK1.1 (PK136)] and 0.25 µg of CD4 (GK1.5) antibodies to target B, CD4+ T, and CD8+ T lymphocytes.
2. Add 100 µL of the antibody mixture to the cell suspension (obtained in step 1.8).
3. Incubate for 25 min on ice.
4. Wash the cells by adding 5 mL of PBS and centrifuging at 400 x g for 5 min at 4 °C, twice.
5. Resuspend the cell pellet in 5 mL of fresh RPMI-1640 medium with 10% FBS and 2.5 µg of DAPI (4',6-diamidino-2-phenylindole).
6. Collect each lymphocyte type separately with flow cytometry using the fluorescence levels of the markers described above. Simultaneously exclude dead cells using the DAPI signals.
   NOTE: Detailed protocols regarding flow cytometry-based cell sorting have been described previously⁴².

3. 3D Quantitative Phase Imaging

1. Keep the sorted lymphocytes on ice throughout the imaging procedures, which should be completed within 5 h (since lymphocyte isolation from the mouse) to avoid cell damage and biochemical alterations.
2. Select a sorted cell type (among B, CD4+ T, and CD8+ T lymphocytes) and dilute the sample (obtained in step 2.6) to 180 cells/µL with RPMI medium for optimal imaging condition (i.e., one cell per single field-of-view).
3. Load 120 μL of the diluted sample into an imaging chamber by slow injection. Thoroughly check on the presence of bubbles in the imaging chamber with the sample. If there are bubbles, carefully remove them, as they will compromise the quality of the measurements.
4. Acquire 3D RI tomograms using a commercial 3D quantitative phase microscope, or holotomography, and its imaging software.
   NOTE: Detailed information about the experimental setup can be found in the original manuscript⁵.
   1. Place a drop of distilled water on top of the objective lens of the microscope.
   2. Place the imaging chamber with the sample on the translation stage of the microscope and adjust its location so that the sample aligns with the objective lens.
   3. Adjust the axial positions of the objective and condenser lenses by clicking Focus and Surface, respectively, on the "Calibration" tab of the "Microscope" perspective of the imaging software.
   4. Align the objective and condenser lenses by clicking Auto Mode. Alternatively, use Scanning Mode and manually adjust the lenses so that the illumination patterns are localized at the central region of the field-of-view.
   5. Return to Normal Mode and adjust the translation stage to locate a cell in the field-of-view.
   6. Find the focal plane by adjusting the axial position of the objective lens. Perfect focusing makes the sample boundary visualized in the screen almost invisible.
   7. Adjust the translation stage to find a location without a cell.
   8. Click Calibrate to measure multiple 2D holograms with varying illumination angles.
   9. Adjust the translation stage to locate a cell at the center of the field-of-view.
   10. Move to the "Acquisition" tab and click 3D Snapshot to measure the holograms of the cell with the identical illumination angles as done in step 3.4.8.
   11. When the acquired data is presented on the "Data Management" panel, right-click on the acquired data and click Process to reconstruct a 3D RI tomogram from the holograms measured in steps 3.4.8 and 3.4.10, using the diffraction tomography algorithm^9,10 implemented in the imaging software.
   12. Repeat steps 3.4.5-3.4.11 to measure more than 100 cells to ensure statistical power for machine learning.
   13. All images that are processed through the step 3.4.11. can be visualized. On the “Data Management” panel, right-click the data and click Open to visualize the data. Click RI Tomogram on the “Data Manager” panel. On the “Preset” tab, click Load and double click “lymphocyte.xml”, which is a predefined transfer function provided by the imaging software to visualize the tomogram according to the 3D RI distributions. 
5. Repeat steps 3.2-3.4 to measure 3D RI tomograms of all lymphocyte subtypes.

4. Quantitative Morphological and Biochemical Feature Extraction from 3D RI Tomograms

1. Place all the tomographic data measured above in a single folder. Split the cell types in the subfolders of this folder. Prepare each tomogram to be a single .mat file.
2. Open Supplementary File 1 - Feature Extraction (written for an image processing software).
3. Edit line 14 to designate the tomogram folder prepared in step 4.1.
4. Edit line 15 to designate a folder to save the extracted feature data.
5. Optionally, edit line 17 to adjust the RI threshold parameters for feature extraction. The default option is 20 RI thresholds from 1.340-1.378, with an increment of 0.002 as described previously⁵.
6. Execute the code. For every tomogram in the dataset, the code calculates five features: surface area (SA), cellular volume (CV), sphericity (SI), protein density (PD), and dry mass (DM), per RI threshold. The detailed algorithms for feature extraction are described elsewhere⁵.
7. In order to monitor the feature extraction, during the execution check on the screen visualizing RI threshold-based cell segmentation.
8. Check on the extracted feature data, as .mat file per tomogram, saved in the folder designated in step 4.4.

5. Supervised Learning and Identification

1. Randomly split the feature data obtained in step 4.8 to training (70%) and test (30%) sets with separate folders.
2. Open Supplementary File 2 - Train (written for an image processing software).
3. Edit line 14 to designate the training set prepared in step 5.1.
4. Edit line 16 to designate a folder to save the trained classifier.
5. Edit line 17 to set a file name for the classifier.
6. Optionally, edit line 19 to select the features for training. The default option, as specified previously⁵, was used to obtain the representative results below.
7. Execute the code. Using the selected features of the training set, the code trains a classifier with the k-nearest neighbor algorithm (k-NN; k = 4) and then saves the classifier (as named in step 5.5) in the folder designated in step 5.4.
8. Check on the screen visualizing the classifier performance and the cross-validation accuracy.
9. Optionally, train multiple classifiers with different feature combinations by repeating steps 5.5-5.7. Then choose the classifier with the highest cross-validation accuracy.
10. Open Supplementary File 3 - Test (written for an image processing software).
11. Edit lines 14-15 to designate the trained classifier to be tested.
12. Edit line 17 to designate the test set prepared in step 5.1.
13. Execute the code. The classifier described above identifies the cell types of the individual lymphocytes in the test set.
14. Check on the screen visualizing the identification performance and the test accuracy.

Results

Figure 1 shows the schematic process of the entire protocol. Using the procedure presented here, we isolated B (n = 149), CD4+ T (n = 95), and CD8+ T (n = 112) lymphocytes. To obtain phase and amplitude information at various angles of illumination, multiple 2D holograms of each lymphocyte were measured by changing the angle of illumination (from -60° to 60°). Typically, 50 holograms can be used to reconstruct a 3D RI tomogram, but the number of 2D hologr...

Discussion

We present a protocol that enables the label-free identification of lymphocyte types exploiting 3D quantitative phase imaging and machine learning. Critical steps of this protocol are quantitative phase imaging and feature selection. For the optimal holographic imaging, the density of cells should be controlled as described above. Mechanical stability of the cells is also important to obtain a precise 3D RI distribution because floating or vibrational cellular motions will disturb hologram measurements upon illumination ...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Mouse	Daehan Biolink	C57BL/6J mice	gender and age-matched, 6 – 8 weeks
Falcon conical centrifuge tube	ThermoFisher Scientific	14-959-53A	15 mL
Phosphate-buffered saline	Sigma-Aldrich	806544-500ML
Ammonium-chloride-potassium lysing buffer	ThermoFisher Scientific	A1049201
RPMI-1640 medium	Sigma-Aldrich	R8758
Fetal bovine serum	ThermoFisher Scientific	10438018
Antibody	BD Biosciences	553140 (RRID:AB_394655)	CD16/32 (clone 2.4G2)
Antibody	BD Biosciences	555275 (RRID:AB_395699)	CD3ε (clone 17A2)
Antibody	Biolegnd	100734 (RRID:AB_2075238)	CD8α (clone 53-6.7)
Antibody	BD Biosciences	557655 (RRID:AB_396770)	CD19 (clone 1D3)
Antibody	BD Biosciences	557683 (RRID:AB_396793)	CD45R/B220 (clone RA3-6B2)
Antibody	BD Biosciences	552878 (RRID:AB_394507)	NK1.1 (clone PK136)
Antibody	eBioscience	11-0041-85 (RRID:AB_464893)	CD4 (clone GK1.5)
DAPI	Roche	10236276001	4,6-diamidino-2-phenylindole
Flow cytometry	BD Biosciences	Aria II or III
Imaging chamber	Tomocube, Inc.	TomoDish
Holotomography	Tomocube, Inc.	HT-1H
Holotomography imaging software	Tomocube, Inc.	TomoStudio
Image professing software	MathWorks	Matlab R2017b