Autofluorescence imaging permits label-free cell type assignment and reveals the dynamic formation of airway secretory cell associated antigen passages (SAPs)

  1. Viral S Shah
  2. Jue Hou
  3. Vladimir Vinarsky
  4. Jiajie Xu
  5. Manalee V Surve
  6. Charles P Lin
  7. Jayaraj Rajagopal

  • Curated by eLife

    eLife logo

    eLife assessment

    This interesting and important methodologic study presents exciting new data identifying approaches to evaluating the cell biology of lung disease. Namely, the ability to identify and track dynamic and coordinated activities of multiple composite cell types in response to experimental interventions. They have developed an interesting label-free approach that collects biologically-encoded autofluorescence of epithelial cells by 2-photon imaging of mouse tracheal explant culture over 2 days. This study has the potential to inform a variety of experimental conditions in lung injury and repair.

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

The specific functional properties of a tissue are distributed amongst its component cell types. The various cells act coherently, as an ensemble, in order to execute a physiologic response. Modern approaches for identifying and dissecting novel physiologic mechanisms would benefit from an ability to identify specific cell types in live tissues that could then be imaged in real time. Current techniques require the use of fluorescent genetic reporters that are not only cumbersome, but which only allow the study of three or four cell types at a time. We report a non-invasive imaging modality that capitalizes on the endogenous autofluorescence signatures of the metabolic cofactors NAD(P)H and FAD. By marrying morphological characteristics with autofluorescence signatures, all seven of the airway epithelial cell types can be distinguished simultaneously in mouse tracheal explants in real time. Furthermore, we find that this methodology for direct cell type-specific identification avoids pitfalls associated with the use of ostensibly cell type-specific markers that are, in fact, altered by clinically relevant physiologic stimuli. Finally, we utilize this methodology to interrogate real-time physiology and identify dynamic secretory cell associated antigen passages (SAPs) that form in response to cholinergic stimulus. The identical process has been well documented in the intestine where the dynamic formation of SAPs and goblet cell associated antigen passages (GAPs) enable luminal antigen sampling. Airway secretory cells with SAPs are frequently juxtaposed to antigen presenting cells, suggesting that airway SAPs, like their intestinal counterparts, not only sample antigen but convey their cargo for immune cell processing.

Article activity feed

  1. Version published to 10.7554/elife.84375 on eLife
  2. eLife

    Author Response

    Reviewer #1 (Public Review):

    The authors optimize a live cell imaging method based on the detection of FAD/NAD(P)H adopted from the fast-growing field of live metabolic imaging. They build upon a method described by KreiB et al 2020 that used metabolic ratio and collagen fiber second harmonic generation imaging. They follow by combining metabolic imaging with morphologic measurements to train a machine-learning model that is able to identify cell types accurately. Upon visualization, authors detected structures hypothesized and then proven to resemble the "goblet cell associated antigen passages" previously studied in intestinal epithelia.

    STRENGTHS

    • The manuscript is succinct, well written, and overall done rigorously.
    • The optimization of the method at multiple levels to the point of identifying both common and rare cell types is impressive.
    • Describes the elegant implementation of a sorely needed method in epithelial biology.
    • Provides an approach to studying the cholinergic response in epithelial cells, a poorly understood phenomenon despite broad clinical use for diagnosis and treatment.

    WEAKNESSES

    A) For what is in large part a methods-development paper, the methods are not explained or shared in a manner that facilitates reproducibility. For example:

    A.1.) The training and validation datasets seem to come from the same sample (or the source is not clearly described). Therefore, it is not clear whether the "96% accuracy" refers to accuracy within the sample measured, or whether it can extrapolate to other samples.

    In order to avoid any confusion, we further clarify that the machine learning training and validation data sets come from the same sample. We had split the total data set into 2 separate subsets for this purpose. This has been laid out in the text as follows:

    “In order to assess the performance of machine learning algorithms designed to distinguish cell types, we divided our data set into training and testing subsets. We utilized 75% of the total cells (154 cells) for machine learning training, leaving 25% (52 cells) for subsequent validation.”

    A.2.) It is unclear whether the model needs to be re-trained within each new sample measured, or if it's applicable to others. This has implications for method adoption by others. Either way is useful but needs to be clarified.

    This is a very interesting point and one that we further clarify in the Discussion noting that in both disease and non-diseased states the model needs to be re-trained in each particular experimental regime.

    A.3.) Code was only listed in a PDF file, which makes reproducing the analysis very cumbersome.

    We hope that all can utilize the code made for this methodology and have uploaded it to a publicly available GitHub account:

    https://github.com/vss11/Label-free-autofluorescence

    B) Whereas the optimization to improve cell type detection is very well described, the implementability of the approach could benefit from exploration (using the data already obtained) of the minimal set of measurements needed to identify cell types. For example, is the FAD/NAD(P)H ratio necessary? Or could just morphologic measurements achieve the same goal?

    This is an excellent point, and we appreciate the Reviewer’s suggestion for this analysis. We have added Figure 3 Supplement 5 where we perform modeling without autofluorescence data. This analysis reveals a dramatic reduction in accuracy with a Matthew’s correlation coefficient ranging from 0.66 to 0.78. This provides additional justification for the use of autofluorescence for cell type identification. Morphologic measurements are not sufficient for cell type identification alone.

    We also have determined the relative contribution of each characteristic to the cell type identification by the Xgboost algorithm in Figure 3 Supplement 4, which shows that autofluorescence signatures are amongst the top contributing characteristics to cell type identification by machine learning.

    C) Whereas the conclusions are overall supported by the data, need small adjustments in some cases:

    C.1.) For example, P3L80: Claims autofluorescence imaging is more specific than "functional markers", however, this is done in the setting of a very specific intervention that massively affects a protein often used as a secretory cell marker (CCSP aka SCGB1A1), which is known to be secreted (and depleted) in secretory cells upon stimulation.

    We agree with the Reviewer that secretory cell identification is a prime example where autofluorescence imaging may be superior to conventional staining, specifically due to the point the Reviewer makes regarding CCSP secretion. We discuss this concept in the Discussion while giving examples of CCSP staining being reduced in asthma, COPD, and smokers. It could be that these cells are missed due to depletion of CCSP. Indeed, we clarify that our methodological approach may be less affected by the loss of the category of specific markers that change with cell state. There are, of course, caveats with utilizing this approach in disease states, and we elaborate on this further below and add this point to the discussion.

    C.2.) Relatedly, it is unclear how the method's accuracy would be affected in conditions that affect redox/metabolic state; the approach may be highly affected in inflammation and injury, for example.

    As suggested by the Reviewer, we re-analyzed the data after Antimycin A + Rotenone and FCCP to determine if autofluorescence ratio is sufficiently different to identify ciliated and secretory cells and included this data in Figure 2 Supplement 1. This is an example where the redox/metabolic state is indeed altered. Though the autofluorescence ratio is affected, it is still useful for cell type identification after intervention as the ciliated and secretory cells have statistically different ratios.

    However, different disease states, particularly infection and inflammation may result in a more profound effect on autofluorescence signatures. For instance, previous work by Dilipkumar et. al, 2019 found changes in autofluorescence over days in repeated measurements in a mouse model of inflammatory bowel disease. Therefore, it is likely that the cell type identification methodology will need to be re-optimized for different experiments and diseased tissues. We include commentary to this effect in the discussion.

    D) The data used to describe "SAPs" is very cursory.

    To further elaborate on our description of SAPs we have included the following:

    1. SAP formation occurs in secretory cells in both stimulated and unstimulated conditions. We performed additional analysis of Figure 4C and determined that SAP formation does occur at baseline prior to stimulation in 9% of secretory cells. Methacholine addition results in 78% of secretory cells forming SAPs (Figure 4 Supplement 1). We have added Figure 5C to demonstrate that SAP formation occurs in the absence of stimulation and is enhanced after methacholine stimulation.

    2. We demonstrate that SAPs can uptake both FITC-dextran and FITC-ovalbumin in Figure 5E, and Figure 5 Supplement 2. We also now show that immune cells (CD11c antigen presenting cells) associate with SAPs containing FITC-dextran and FITC-ovalbumin in Figure 5E and Figure 5 Supplement 2. We have expanded the Discussion of SAPs.

    3. We now show 3 video examples and an XZ optical cross section of ALI that demonstrate uptake and secretion of FITC-dextran in Figure 5 Supplemental Videos 1-3 and Figure 5 Supplement 1.

    D.1.) Unclear if FITC dextran uptake occurs in other cells too, or in secretory cells prior to methacholine stimulation, or induced nonspecifically due to epithelia manipulation. Secretory and goblet cells are very sensitive to stimulation and often considered minimal, for example, see the paper by Abdullah et al DOI:10.1007/978-1-61779-513-8_16 in which extreme care had to be applied to prevent any secretion at all.

    Our autofluorescence methodology revealed the formation of “voids” of autofluorescence forming in secretory cells and we focused our experiments on this phenomenon. Based on the reviewer question, we generated Figure 5C to better characterize SAP formation. Figure 5C illustrates that SAP formation occurs in both unstimulated and methacholine stimulated conditions, but is dramatically increased following methacholine stimulation. This is analogous to the behavior of GAPs in the intestine (Knoop et al., 2015). Furthermore, we have reanalyzed Figure 4C to identify SAPs prior to stimulation and found that these structures are present in 9% of secretory cells. After methacholine stimulation this percentage increases to 78%.

    D.2.) A single image is provided for the SAP timeline (Figure 5C), which appears to be the same cell shown in the supplementary video.

    We now provide numerous example videos and optical XZ cross section of ALI demonstrating SAP uptake and secretion in Supplementary Videos 1-3 and Figure 5 Supplement 1.

    IMPACT AND UTILITY

    This is well-done work with high potential for widespread adoption within the epithelial biology community, particularly if the methods and code are shared in better detail.

    We indeed hope that this methodology can be utilized by others. We have posted analysis code, raw data, MATLAB algorithm, and other necessary files onto a publicly available GitHub link. https://github.com/vss11/Label-free-autofluorescence

    Reviewer #2 (Public Review):

    Shah and colleagues tackle a significant impediment to exploiting tissue culture systems that enable prospective ex vivo experimentation in real-time. Namely, the ability to identify and track dynamic and coordinated activities of multiple composite cell types in response to experimental perturbations. They develop a clever label-free approach that collects biologically-encoded autofluorescence of epithelial cells by 2-photon imaging of mouse tracheal explant culture over 2 days. They report the ability to distinguish 7 cell types simultaneously, including rare ones, by developing a machine-learning approach using a combination of fluorescence and cytologic features. Their algorithm demonstrates high accuracy by Mathew's Correlation Coefficient when applied to a test set. Lastly, they show the ability of their approach to visualize the dynamic uptake and expulsion of fluorescently-tagged dextran by individual secretory cells. Overall, the results are intriguing and may be very useful for specific applications.

    We thank the reviewers for their assessment and indeed hope that the methodology is useful and the discovery of the dynamics of SAP formation have important implications for airway mucosal immunology.

  3. eLife

    eLife assessment

    This interesting and important methodologic study presents exciting new data identifying approaches to evaluating the cell biology of lung disease. Namely, the ability to identify and track dynamic and coordinated activities of multiple composite cell types in response to experimental interventions. They have developed an interesting label-free approach that collects biologically-encoded autofluorescence of epithelial cells by 2-photon imaging of mouse tracheal explant culture over 2 days. This study has the potential to inform a variety of experimental conditions in lung injury and repair.

  4. eLife

    Reviewer #1 (Public Review):

    The authors optimize a live cell imaging method based on the detection of FAD/NAD(P)H adopted from the fast-growing field of live metabolic imaging. They build upon a method described by KreiB et al 2020 that used metabolic ratio and collagen fiber second harmonic generation imaging. They follow by combining metabolic imaging with morphologic measurements to train a machine-learning model that is able to identify cell types accurately. Upon visualization, authors detected structures hypothesized and then proven to resemble the "goblet cell associated antigen passages" previously studied in intestinal epithelia.

    STRENGTHS
    - The manuscript is succinct, well written, and overall done rigorously.
    - The optimization of the method at multiple levels to the point of identifying both common and rare cell types is impressive.
    - Describes the elegant implementation of a sorely needed method in epithelial biology.
    - Provides an approach to studying the cholinergic response in epithelial cells, a poorly understood phenomenon despite broad clinical use for diagnosis and treatment.

    WEAKNESSES
    A) For what is in large part a methods-development paper, the methods are not explained or shared in a manner that facilitates reproducibility. For example:
    A.1.) The training and validation datasets seem to come from the same sample (or the source is not clearly described). Therefore, it is not clear whether the "96% accuracy" refers to accuracy within the sample measured, or whether it can extrapolate to other samples.
    A.2.) It is unclear whether the model needs to be re-trained within each new sample measured, or if it's applicable to others. This has implications for method adoption by others. Either way is useful but needs to be clarified.
    A.3.) Code was only listed in a PDF file, which makes reproducing the analysis very cumbersome.

    B) Whereas the optimization to improve cell type detection is very well described, the implementability of the approach could benefit from exploration (using the data already obtained) of the minimal set of measurements needed to identify cell types. For example, is the FAD/NAD(P)H ratio necessary? Or could just morphologic measurements achieve the same goal?

    C) Whereas the conclusions are overall supported by the data, need small adjustments in some cases:
    C.1.) For example, P3L80: Claims autofluorescence imaging is more specific than "functional markers", however, this is done in the setting of a very specific intervention that massively affects a protein often used as a secretory cell marker (CCSP aka SCGB1A1), which is known to be secreted (and depleted) in secretory cells upon stimulation.
    C.2.) Relatedly, it is unclear how the method's accuracy would be affected in conditions that affect redox/metabolic state; the approach may be highly affected in inflammation and injury, for example.

    D) The data used to describe "SAPs" is very cursory.
    D.1.) Unclear if FITC dextran uptake occurs in other cells too, or in secretory cells prior to methacholine stimulation, or induced nonspecifically due to epithelia manipulation. Secretory and goblet cells are very sensitive to stimulation and often considered minimal, for example, see the paper by Abdullah et al DOI:10.1007/978-1-61779-513-8_16 in which extreme care had to be applied to prevent any secretion at all.
    D.2.) A single image is provided for the SAP timeline (Figure 5C), which appears to be the same cell shown in the supplementary video.

    IMPACT AND UTILITY
    This is well-done work with high potential for widespread adoption within the epithelial biology community, particularly if the methods and code are shared in better detail.

  5. eLife

    Reviewer #2 (Public Review):

    Shah and colleagues tackle a significant impediment to exploiting tissue culture systems that enable prospective ex vivo experimentation in real-time. Namely, the ability to identify and track dynamic and coordinated activities of multiple composite cell types in response to experimental perturbations. They develop a clever label-free approach that collects biologically-encoded autofluorescence of epithelial cells by 2-photon imaging of mouse tracheal explant culture over 2 days. They report the ability to distinguish 7 cell types simultaneously, including rare ones, by developing a machine-learning approach using a combination of fluorescence and cytologic features. Their algorithm demonstrates high accuracy by Mathew's Correlation Coefficient when applied to a test set. Lastly, they show the ability of their approach to visualize the dynamic uptake and expulsion of fluorescently-tagged dextran by individual secretory cells. Overall, the results are intriguing and may be very useful for specific applications.

  6. Version published to 10.1101/2022.11.01.514675v1 on bioRxiv