2017年[ 技術開発研究助成 (開発研究) ] 成果報告 ： 年報30号補刷
非標識マルチモーダル顕微鏡法を用い細胞状態計測法の開発 Label-free multimodal microscopy for cellular state investigation
概要
1. Introduction
Many techniques designed to observe cellular state or responses to known stimuli rely on the use of chemical labels, which are employed to target specific molecules which are thought to be involved in the studied process. These methods consist for instance in staining histologic tissue sections to enhance contrast, or the insertion of fluorescent dyes into cells to observe the spatio-temporal features of molecules of interest.
We are focusing in this project on immune cells, and in particular macrophages. In the context of immunology, there is often a need for studies based on large populations, due to the wide diversity of cells and interaction mechanisms involved in the immunological responses. For this reason, analysis methods such as flow cytometry are often employed, providing analyzes based on large samples for strong statistical significance. Furthermore, methods based on fluorescent dyes require that a known molecule of interest and a specific marker which bonds to it are necessary to perform the analysis. In the context of immunology, for example in the case of macrophage stimulation, there are not many markers known to efficiently discriminate activated and resting cells. However, as immune cells are known to communicate through secreted molecules such as cytokines, it is often more efficient to assess the activation of cells by detecting the presence of specific cytokines in the extracellular medium [1]. This is classically performed through enzyme-linked immunosorbant assay (ELISA), which selectively bonds to the cytokine with antibodies. While this approach is very sensitive, it only provides an indication of activation at population level, without indications of the state of individual cells.
While these different methods provide very sensitive ways to assess the activation state of cells, they require knowing and selecting target molecules to analyze beforehand. Furthermore, most of these techniques do not make it possible to relate the results to individual cells that could be targeted for further analysis.
The purpose of this project is to evaluate the possibility of employing label-free microscopy to assess the activation state of macrophage cells in situ and non-invasively, enabling further analysis on the measured cells which can be maintained in an unperturbed environment. We also focus on high-throughput data acquisition, to make it possible to measure a large amount of cells rapidly enough to retrieve statistically significant results. We employ for this purpose a multimodal system comprising several imaging methods that are described below.
Quantitative phase microscopy (QPM) is a full-field imaging method which provides contrast based on the optical density of the specimen. Contrarily to more well-known techniques such as Zernike’s phase contrast or differential interference contrast, QPM provides quantitative information about the local intracellular content, and in particular proteins [2]. We employ here a QPM technique based on digital holography, which provides one-shot recovery of the phase images based on laser illumination.
Raman spectroscopy is an optical method in which the resonant response of the intracellular chemical content to an excitation beam is recorded. As each molecule possesses specific vibrational modes, it is possible to derive an optical signature that provides insight about the molecular content in a non-invasive way. While this approach can provides very specific information, it has the disadvantage of being relatively slow, making it not very suitable for high-throughput imaging. We employ here a method we developed previously, denoted as hybrid scanning [3], that can provides a relevant Raman spectrum of the cell based on point measurements, which can be acquired much faster.
While fluorescence commonly requires the insertion of labels in order to enable imaging, it is also possible to image the inherent autofluorescence of biomolecules. It is known that activated macrophages secrete reactive oxygen species (ROS) such as nitric oxide which has a cytotoxic effect against pathogens [4]. As ROS have a strong autofluorescence, imaging this response provides an additional contrast for analysis, which is based on the concentration of molecules naturally occurring during activation.
2. Technical developments
The measurement setup we developed is based on an existing system that had been developed to enable simultaneous measurement of Raman imaging and QPM [2, 3]. The system has been then modified to include a standard wide-field fluorescence system (see Fig.1), which can be used to image the same field of view as with the Raman and QPM modalities by employing flip mirrors that insert the illumination and detection elements of an epi-fluorescence system. This approach provides a simple way to measure the autofluorescence in parallel to the two other channels that can be measured simultaneously, as shown in Fig.1.
The measurement procedure has also been improved in terms of speed and responsiveness. In order to retrieve the phase images, a reconstruction algorithm must be applied to the measurements, which leads to typical frame rates in the range of 5 frames per second for live reconstructions. We developed a new reconstruction code which relies on graphical processing units, that can reach reconstruction rates up to 50 frames per second, enabling real-time monitoring during measurements.
（注:図/PDFに記載）
After measurement, the images of both QPM and autofluorescence can be combined through a registration procedure for further analysis (see Fig.2). One can see that the two modalities show rather different contrasts, with QPM (green) highlighting the overall cell shape, with nucleoli and some cytoplasmic structure visible. On the other hand, autofluorescence (red) provides contrast mainly in the cytoplasm, with more structure within cells. The two images clearly possess complementary information.
Based on the recorded data, we then developed a protocol to analyze the registered images and segment the cells to extract morphological parameters for further analysis. This has been performed through the CellProfiler program, which is distributed as an open source code [5].
Several parameters can then be extracted from the cell images based on the cellular regions, either from the quantitative values from QPM images, or from autofluorescence images, whose intensity is related to ROS concentration. On the other hand, Raman spectra can be processed to be representative of the whole cell content and comparable for further analysis [6].
（注:図/PDFに記載）
3. Validations
We first ensure that the experimental setup can detect the elevation of ROS levels upon standard stimulations. To this end, we measure the fluorescence from cells after exposure to lipopolysaccharide (LPS), which simulates bacterial infection. We compare the autofluorescence signal with the fluorescence from the CellROX dye, which measures the oxidative stress level.
As shown in Fig.3, the overall signal is significantly higher in stimulated cells, as expected from the ROS concentration increase under activation. One can note that autofluorescence has a larger difference, probably due to the larger amount of endogenously emitting molecules. However, this difference is not sufficient to provide a satisfactory classification, as shown by the large variation.
（注:図/PDFに記載）
We also validated the algorithmic approach, by first testing various methods to segment the cells. Interestingly, the QPM images provide much reliable results than fluorescence, thanks to their signal representing the overall shape of the cell, and the possibility of largely compensating for background artifacts in phase images, while autofluorescence images led to a large amount of sub-divisions. The result of segmentation is shown in Fig.2, where white outlines represent the selected regions.
We also tested different approach to develop the statistical model based on the extracted parameters from label-free images and signals, ranging from linear methods (linear discriminant analysis, logistic regression) to non-linear ones (polynomial regression, support vector machine) [7]. We found that linear classifiers combined with assisted variable reduction to reduce the amount of employed parameters provided the most stable results.
4. Single cell classification
We then applied the measurement and analysis protocols described above to develop a statistical model of stimulated macrophage cells which could then reliably predict the activation state of single cells. To prepare and test the model, we measured Raw264.7 cells, which is a macrophage-like mouse cell line derived from lung-residing macrophages. We measured cells plated on several dishes to account for culture differences, and stimulated them with LPS to simulate a bacterial infection. We then extracted a statistical model based on hundreds of cells explaining the difference between stimulated and control dishes. To validate our approach, we then applied our statistical model to other batches of cells, which were measured in similar conditions, but on different days throughout weeks following the initial measurements, to account for possible differences in cell health conditions, or calibration issues in the experimental system. The results show that we can consistently reach around 85?90% prediction accuracy between the expected cell activation state and the presence of the stimulus. Furthermore, the results are very consistent between cross-validated training and testing data, a strong indication of a stable model. Practically, the prediction yields a probability of being activated for each single cell, where the difference between control and LPS dishes can clearly be seen from the cells scores distributions within dishes (see Fig.4A). The results also lead to population densities (see Fig.4B), where the two main groups can be clearly identified. These results are also confirmed with ELISA assays performed on the supernatant of the culture medium after measurements, where TNF-α levels, a secreted cytokine known to be released by activated macrophages, show that the cells, as a population, are activated in LPS dishes as shown in Fig.4C.
It is also possible to employ the spectral data to study the molecular content of cells, as shown in Fig.5A with a typical cell spectrum, and to employ the Raman spectra for developing a model enabling the recognition of the activation state of macrophage cells, even though the use of continuous data like a spectrum requires more steps than for discrete parameters. This can be seen in Fig.5B, where cells from control and stimulated dishes are clearly separated in different populations through the analysis of spectral features. Interestingly, the morphological and spectral features are highly complementary. While one can imagine that morphological features can be dependent on cells conditions, Raman measurements are rather stable in regards to cell health, but can depend on the substrate quality, as shown in Fig.5C. Combining the two types of information can then contribute to derive more stable predictors.
（注:図/PDFに記載）
5. Discussion
The approach proposed here enables the possibility of assessing a complex biological behavior such as macrophage activation based on label-free indicators, so that cells can be observed directly in culture without any preliminary preparation protocol.
As the technique is based on imaging, it is possible to extract the parameters and derive the estimator for each cell in the field of view, leading to high-throughput single-cell estimators. Based on a combination of morphological, structural and chemical parameters derived from QPM and autofluorescence, it is possible to derive a statistical model which can reliably predict the activation state of cells measured several days or weeks apart, despite the fact that the changes induced by stimulation with LPS on macrophages cells are expected to be small, as it consists in a reversible change within a cell type known for its plasticity.
（注:図/PDFに記載）
Interestingly, as the method is purely based on label-free measurements, so that all signals are generated by endogenous content, it implies that the same approach can be applied to other targets, provided that training data is available. As the statistical model is developed based on existing measurements, another model could readily be developed to suit the objective target for further predictions.
Acknowledgments
This research has been supported by the Nakatani Foundation for Advancement of Measuring Technologies in Biomedical Engineering through a grant for research development (31st edition), and by the Japan Society for the Promotion of Science (JSPS) World Premier International Research Center Initiative Funding Program.
Many techniques designed to observe cellular state or responses to known stimuli rely on the use of chemical labels, which are employed to target specific molecules which are thought to be involved in the studied process. These methods consist for instance in staining histologic tissue sections to enhance contrast, or the insertion of fluorescent dyes into cells to observe the spatio-temporal features of molecules of interest.
We are focusing in this project on immune cells, and in particular macrophages. In the context of immunology, there is often a need for studies based on large populations, due to the wide diversity of cells and interaction mechanisms involved in the immunological responses. For this reason, analysis methods such as flow cytometry are often employed, providing analyzes based on large samples for strong statistical significance. Furthermore, methods based on fluorescent dyes require that a known molecule of interest and a specific marker which bonds to it are necessary to perform the analysis. In the context of immunology, for example in the case of macrophage stimulation, there are not many markers known to efficiently discriminate activated and resting cells. However, as immune cells are known to communicate through secreted molecules such as cytokines, it is often more efficient to assess the activation of cells by detecting the presence of specific cytokines in the extracellular medium [1]. This is classically performed through enzyme-linked immunosorbant assay (ELISA), which selectively bonds to the cytokine with antibodies. While this approach is very sensitive, it only provides an indication of activation at population level, without indications of the state of individual cells.
While these different methods provide very sensitive ways to assess the activation state of cells, they require knowing and selecting target molecules to analyze beforehand. Furthermore, most of these techniques do not make it possible to relate the results to individual cells that could be targeted for further analysis.
The purpose of this project is to evaluate the possibility of employing label-free microscopy to assess the activation state of macrophage cells in situ and non-invasively, enabling further analysis on the measured cells which can be maintained in an unperturbed environment. We also focus on high-throughput data acquisition, to make it possible to measure a large amount of cells rapidly enough to retrieve statistically significant results. We employ for this purpose a multimodal system comprising several imaging methods that are described below.
Quantitative phase microscopy (QPM) is a full-field imaging method which provides contrast based on the optical density of the specimen. Contrarily to more well-known techniques such as Zernike’s phase contrast or differential interference contrast, QPM provides quantitative information about the local intracellular content, and in particular proteins [2]. We employ here a QPM technique based on digital holography, which provides one-shot recovery of the phase images based on laser illumination.
Raman spectroscopy is an optical method in which the resonant response of the intracellular chemical content to an excitation beam is recorded. As each molecule possesses specific vibrational modes, it is possible to derive an optical signature that provides insight about the molecular content in a non-invasive way. While this approach can provides very specific information, it has the disadvantage of being relatively slow, making it not very suitable for high-throughput imaging. We employ here a method we developed previously, denoted as hybrid scanning [3], that can provides a relevant Raman spectrum of the cell based on point measurements, which can be acquired much faster.
While fluorescence commonly requires the insertion of labels in order to enable imaging, it is also possible to image the inherent autofluorescence of biomolecules. It is known that activated macrophages secrete reactive oxygen species (ROS) such as nitric oxide which has a cytotoxic effect against pathogens [4]. As ROS have a strong autofluorescence, imaging this response provides an additional contrast for analysis, which is based on the concentration of molecules naturally occurring during activation.
2. Technical developments
The measurement setup we developed is based on an existing system that had been developed to enable simultaneous measurement of Raman imaging and QPM [2, 3]. The system has been then modified to include a standard wide-field fluorescence system (see Fig.1), which can be used to image the same field of view as with the Raman and QPM modalities by employing flip mirrors that insert the illumination and detection elements of an epi-fluorescence system. This approach provides a simple way to measure the autofluorescence in parallel to the two other channels that can be measured simultaneously, as shown in Fig.1.
The measurement procedure has also been improved in terms of speed and responsiveness. In order to retrieve the phase images, a reconstruction algorithm must be applied to the measurements, which leads to typical frame rates in the range of 5 frames per second for live reconstructions. We developed a new reconstruction code which relies on graphical processing units, that can reach reconstruction rates up to 50 frames per second, enabling real-time monitoring during measurements.
（注:図/PDFに記載）
After measurement, the images of both QPM and autofluorescence can be combined through a registration procedure for further analysis (see Fig.2). One can see that the two modalities show rather different contrasts, with QPM (green) highlighting the overall cell shape, with nucleoli and some cytoplasmic structure visible. On the other hand, autofluorescence (red) provides contrast mainly in the cytoplasm, with more structure within cells. The two images clearly possess complementary information.
Based on the recorded data, we then developed a protocol to analyze the registered images and segment the cells to extract morphological parameters for further analysis. This has been performed through the CellProfiler program, which is distributed as an open source code [5].
Several parameters can then be extracted from the cell images based on the cellular regions, either from the quantitative values from QPM images, or from autofluorescence images, whose intensity is related to ROS concentration. On the other hand, Raman spectra can be processed to be representative of the whole cell content and comparable for further analysis [6].
（注:図/PDFに記載）
3. Validations
We first ensure that the experimental setup can detect the elevation of ROS levels upon standard stimulations. To this end, we measure the fluorescence from cells after exposure to lipopolysaccharide (LPS), which simulates bacterial infection. We compare the autofluorescence signal with the fluorescence from the CellROX dye, which measures the oxidative stress level.
As shown in Fig.3, the overall signal is significantly higher in stimulated cells, as expected from the ROS concentration increase under activation. One can note that autofluorescence has a larger difference, probably due to the larger amount of endogenously emitting molecules. However, this difference is not sufficient to provide a satisfactory classification, as shown by the large variation.
（注:図/PDFに記載）
We also validated the algorithmic approach, by first testing various methods to segment the cells. Interestingly, the QPM images provide much reliable results than fluorescence, thanks to their signal representing the overall shape of the cell, and the possibility of largely compensating for background artifacts in phase images, while autofluorescence images led to a large amount of sub-divisions. The result of segmentation is shown in Fig.2, where white outlines represent the selected regions.
We also tested different approach to develop the statistical model based on the extracted parameters from label-free images and signals, ranging from linear methods (linear discriminant analysis, logistic regression) to non-linear ones (polynomial regression, support vector machine) [7]. We found that linear classifiers combined with assisted variable reduction to reduce the amount of employed parameters provided the most stable results.
4. Single cell classification
We then applied the measurement and analysis protocols described above to develop a statistical model of stimulated macrophage cells which could then reliably predict the activation state of single cells. To prepare and test the model, we measured Raw264.7 cells, which is a macrophage-like mouse cell line derived from lung-residing macrophages. We measured cells plated on several dishes to account for culture differences, and stimulated them with LPS to simulate a bacterial infection. We then extracted a statistical model based on hundreds of cells explaining the difference between stimulated and control dishes. To validate our approach, we then applied our statistical model to other batches of cells, which were measured in similar conditions, but on different days throughout weeks following the initial measurements, to account for possible differences in cell health conditions, or calibration issues in the experimental system. The results show that we can consistently reach around 85?90% prediction accuracy between the expected cell activation state and the presence of the stimulus. Furthermore, the results are very consistent between cross-validated training and testing data, a strong indication of a stable model. Practically, the prediction yields a probability of being activated for each single cell, where the difference between control and LPS dishes can clearly be seen from the cells scores distributions within dishes (see Fig.4A). The results also lead to population densities (see Fig.4B), where the two main groups can be clearly identified. These results are also confirmed with ELISA assays performed on the supernatant of the culture medium after measurements, where TNF-α levels, a secreted cytokine known to be released by activated macrophages, show that the cells, as a population, are activated in LPS dishes as shown in Fig.4C.
It is also possible to employ the spectral data to study the molecular content of cells, as shown in Fig.5A with a typical cell spectrum, and to employ the Raman spectra for developing a model enabling the recognition of the activation state of macrophage cells, even though the use of continuous data like a spectrum requires more steps than for discrete parameters. This can be seen in Fig.5B, where cells from control and stimulated dishes are clearly separated in different populations through the analysis of spectral features. Interestingly, the morphological and spectral features are highly complementary. While one can imagine that morphological features can be dependent on cells conditions, Raman measurements are rather stable in regards to cell health, but can depend on the substrate quality, as shown in Fig.5C. Combining the two types of information can then contribute to derive more stable predictors.
（注:図/PDFに記載）
5. Discussion
The approach proposed here enables the possibility of assessing a complex biological behavior such as macrophage activation based on label-free indicators, so that cells can be observed directly in culture without any preliminary preparation protocol.
As the technique is based on imaging, it is possible to extract the parameters and derive the estimator for each cell in the field of view, leading to high-throughput single-cell estimators. Based on a combination of morphological, structural and chemical parameters derived from QPM and autofluorescence, it is possible to derive a statistical model which can reliably predict the activation state of cells measured several days or weeks apart, despite the fact that the changes induced by stimulation with LPS on macrophages cells are expected to be small, as it consists in a reversible change within a cell type known for its plasticity.
（注:図/PDFに記載）
Interestingly, as the method is purely based on label-free measurements, so that all signals are generated by endogenous content, it implies that the same approach can be applied to other targets, provided that training data is available. As the statistical model is developed based on existing measurements, another model could readily be developed to suit the objective target for further predictions.
Acknowledgments
This research has been supported by the Nakatani Foundation for Advancement of Measuring Technologies in Biomedical Engineering through a grant for research development (31st edition), and by the Japan Society for the Promotion of Science (JSPS) World Premier International Research Center Initiative Funding Program.