Human lung adenocarcinoma displays significant cell and nuclear volume dysregulation that can be measured in 3D by exploiting machine learning

Human lung alveoli are structured air sacs with lumens which are surrounded by an epithelium. This epithelial layer is composed alveolar type 1 (AT1) and type 2 (AT2) cells (Fig 1A) and lies on top of a thin stromal layer. AT2 cells fulfill the critical functions of secreting surfactants that are released into the lamina of the alveolus from multilamellar bodies, and are the resident stem cells of the alveolus [42, 44, 45, 47, 48, 50, 51]. AT1 cells differentiate from dividing AT2 cells that undergo a process of shaping and flattening to a 1 μm thickness and loss of surfactant protein expression [45]. LA is believed to result from an oncogenic event in dividing AT2 cells in mice, although evidence for this in humans is limited [44]. LA cells are known to overexpress cytokeratin 7 before they undergo additional transformations [52].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Human lung adenocarcinoma is characterized by dysregulation of cell and nuclear size.

63X images shown. A., schematized alveolus, related to B., showing AT2 cells (marker: proSPC, green), presumptive wide and flat AT1 cells, non-CK7-staining macrophage, M, and stroma, S. Epithelial cells express marker CK7, but weakly in AT2 cell bodies. B. and C., 5 μm z-projections of representative normal distant control alveoli, B., and representative EGFR+ LA tumor, AT2 cells marked by proSPC (green), C. Cytokeratin 7, CK7 (red); overexpressed in tumor cells. DNA stained by TO-PRO-3 (cyan). White arrows: thin cellular processes (shown in 1E, below). Right, full-stack max-intensity projections from each axis. Scale bars, 20 μm. Row D., 3D segmentation approach for tumor cells. Left, a high-density confocal stack often referenced during early development of our machine-learning algorithm. Middle, Machine Learning from CK7 staining predicts cell boundaries. Right, Commercial Imaris software fits cell boundaries, with rainbow coloring showing individually segmented cell surfaces. High-coverage case shown. E. Left, 3D rendering of AT2 cells (proSPC) showing thin cellular processes; right, processes shown with AT2 cell body segmentation (gray). White box shows perspective scale, 2 μm per small tick mark. Yellow arrow: AT2 cell body; white arrow: cellular process.

https://doi.org/10.1371/journal.pone.0274091.g001

We first asked whether it was possible to distinguish LA cells from their proposed AT2 progenitors in whole-mount tumor sections and to reconstruct three dimensional morphologies. In doing so we reasoned it would be possible to identify alterations in cell and organelle dimensions in situ in the context of local tissue organization. We performed whole-mount immunofluorescence (IF) confocal imaging of formaldehyde-fixed and paraffin-embedded (FFPE) [53] samples that comprised 150 μm-sections of pre-identified normal and LA patient specimens. FFPE is well-documented to undergo minor isotropic shrinkage over the length scale order of 1–100 μm relevant here, which depend on a given lab’s procedures [54]. However, it preserves cell and nuclear features and DNA content very well, while providing a route to deep image characterization of human samples that are already phenotyped, genotyped, and with well-defined treatment history [55]. To identify cell types in situ, we immunostained using antibodies specific to the unprocessed N-terminal region of prosurfactant protein C (proSPC), which is highly expressed in the cytoplasm of AT2 cells, and cytokeratin 7, which labels epithelial and LA cells with high sensitivity; it does not stain underlying stroma [42, 43, 48, 52, 56, 57]. To stain DNA for nucleus segmentation we used TO-PRO-3. For DNA content measurements we included stromal cell nuclei to establish normal baseline levels of DNA. At high magnification our image stack was typically ~135 x 135 x 60 μm [58].

Representative IF images of normal and LA specimens (Fig 1B and 1C) illustrate that delineation of AT2 and CK7-overexpressing tumor cells was straightforward and facilitated observation of novel cellular morphologies and changes to the cell size and shape. AT2 cells were easily identified as bright-green fluorescing cells (proSPC) mutually exclusive with the red channel in their cytoplasm (CK7), but bracketed by brightly red-fluorescing alveolar walls. These cells also have intense TO-PRO-3 DNA stains typical of cell nuclei (AT1 cells, Fig 1A shows a schematic). We observed AT2 cell bodies filled with dense green-fluorescing puncta, which mark multilamellar vesicles containing proSPC as expected for these secretory cells. Surprisingly, we also noticed that this protein marker extended into an interconnected network of thin “processes”, which to our knowledge have not previously been described (Fig 1B). In pre-identified tumor specimens, tumor cells were visibly obvious, presenting high intensity of red fluorescence in cytoplasm (CK7). Additionally, LA epithelial cells showed a dramatic enlargement of their cell bodies and nuclei (Fig 1C). LA cell nuclei are as large as cell bodies of normal AT2 cells and, at 63X magnification, display proSPC processes confined to the surface boundary of cell-cell junctions (Fig 1C, sidebar shows side-perspectives of channel 1). Tumor cell processes do not connect to cells classified as AT2, but in rare cases, tumor cell processes did connect directly to neighboring cells of an AT2-like morphology (S4B Video, punctate cells and S5 Video, yellow arrows). Concomitant with the enlargement of tumor nuclei, DNA organization often changes, becoming more spread-out and granular or cavitated appearance relative to DNA staining in stromal neighbors (Fig 1C). CK7 staining reveals cell boundaries of the epithelial cells and hints at a broadened range of cell volumes relative to normal AT2 cells (Fig 1C). Highly elongated stromal nuclei highly reminiscent of smooth muscle layers but of unknown type were also frequently observed (Fig 1D, “S”, and S4B Video, yellow circle).

To determine quantitatively whether these LA epithelial cells and their nuclei had enlarged we had to develop computational strategies for image segmentation. Due to challenges with immunostaining traditional cell boundary markers we ultimately decided to use CK7 stain for epithelial cell body segmentation. However the CK7 staining pattern precluded use of classical watershed approaches. To overcome this hurdle and automate cell body segmentation for hundreds of objects per stack, we trained a UNet-based convolutional neural network with a set of 40 optimized manual cell body annotations (S1 Fig). The resulting spaghetti model is a plot of the network’s confidence in a boundary (Fig 1D, middle, and S1 and S2 Figs), and represents 3D scatter data that can be fit by a surface model using Bitplane Imaris^™ software (S3 Fig). Segmentation of AT2 cells from normal tissue was performed in parallel using straightforward thresholding approaches, with a comparable level of accuracy, as assessed by agreement with proSPC or CK7 data (S1D, S1E and S2–S4 Figs). Staining of AT2 cells also revealed thin cellular processes (Fig 1E), discussed later in the manuscript. Segmentation of nuclei was carried out using DNA stain and the cell surface models, correcting for variations in stack brightness (S5 Fig). Due to cell crowding in the tumor, we also manually corrected small clippings to construct an accurate nuclear model (S4 Fig). Notably, nuclear volume segmentation usually required filling cavities present in binary representations of nuclear volumes, highlighting that large regions of oversized tumor cell are DNA-poor (S5 Fig).

Initial 3D surfaces inevitably contained some inaccurate cell models and thus every cell and nucleus was subjected to thorough screening for qualitative agreement with CK7 and DNA data in both 2D and 3D (S3 Fig and Methods). The resulting quality-screened data are provided in S1 Data, tab B., and all quantitative data shown and discussed below are for cells passing this quality assessment screening. In general, the machine learning approach was able to recognize a cell boundary to a similar degree as the authors, and cells failing QC are indicative of poor CK7 staining. Thus, our dataset includes normal lung and LA lung epithelia that the human eye can clearly detect in the microscope and with clear boundaries that stain brightly with our IF protocol. An unfortunate limitation of our data is that it does not include small neoplasms that stain with some proSPC, but not with CK7. An overview of cell models with good cell coverage are shown in S1–S5 Videos; S4B and S5 Videos emphasize that small neoplasms and lesions stain poorly with CK7 and are thus not included in our data, but may be relevant to tumor cell development (green channels shown late in video and S5 Video, yellow arrows).

Quantitation of cell and organelle dimensions in KRAS+ and EGFR+ lung tumors

It is commonly observed that proliferating cells in tumors display a high nucleus:cytoplasm (N:C) volume ratio [13, 14]. However, observations of cell volume dysregulation are less frequently reported. Therefore, we wondered whether reconstructed LA cells, filtered for near-euploid DNA signal, would display only alterations of N:C ratio or whether they might show bona fide increases in cell and nucleus volume (Fig 2A). We found that for both genotypes of LA, epithelial cells had cell volumes and nucleus volume that drastically differed from normal AT2 cells (Kruskal-Wallis and ANOVA, p < 10⁻⁴, Fig 2B–2D, S6–S8 Figs). AT2 cell and nuclear volume data with overall mean ± SD of 682.3 ± 194.2 fit to a 2-Gaussian mixture model to recover a first peak of cell volume ± σ of 582 ± 127 fL (n = 802, χ²/df = 1.265 x 10⁻⁸) and nuclear volume of 202 ± 34.2 fL (n = 802, χ²/df = 9.285 x 10⁻⁸). In contrast epithelial cells across all stages of LAs have mean ± SD cell volumes of 1334 ± 1340 fL (n = 4133), and nuclear volumes of 553.3 ± 378.6 fL (n = 4082). Considering only near-euploid cells, defined as 1.6–4.4n (n = haploid genome), gives cell volumes of 1086 ± 755.6 fL (n = 2683) and nuclear volumes of 462.3 ± 222.2 fL (n = 2671, S6 and S7 Figs). However, inspection of the cell volume distributions and outliers in the data reveals that eliminating cells >4.4n results in elimination of outliers in the long tail >7 pL and loss of the long tail in the distribution (Fig 2B and S6–S8 Figs). Thus, in the absence of ploidies > 4.4n, size dysregulation, defined as standard deviation (S.D.), is 5.9-fold higher for cell volume and 6.5-fold higher for nuclear volume in diseased versus normal (Fig 2C and 2E). Surprisingly, alterations of N:C ratio were much more modest (Fig 2E and S8 Fig) compared to the extent of cell and nucleus size alteration (Fig 2C and 2D). Importantly, polyploidy contributes to size dysregulation by a 77.6% increase in the S.D. of tumor cell volume relative to near-euploid tumor cells (Fig 2F). However, we still observe an 86.6% increase in mean volume of near-euploid tumor cells relative to 2n AT2 wild-type cells. Thus, polyploidy is not necessary to observe a nearly 2-fold cell-size increase in human LA.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Quantitation of cell and nucleus dimensions in EGFR+ and KRAS+ lung tumors.

A., Schematic: potential outcomes for cell and nucleus scaling in cancerous cells, hypothesis 2 is that LA retains nuclear scaling function. B., histograms of cell volumes of near-euploid (1.6–4.4n) cells (N = 21, n = 2683) as a function of EGFR+ (N = 10, n = 1411) or KRAS+ (N = 11, n = 1260) genotype and stage. Green, cell volumes for tumor cells within the boundaries of the 2n normal AT2 cells; Red, cell volumes outside the Gaussian distribution of 2n population of Normal AT2 cell volumes. C., violin plots for near-euploid cells show very significantly abnormal scaling and significant changes in distributions between stages. 1–2 and 2–3. Colors alternating for visibility, normal in grey. Kruskal-Wallis ANOVA, p<1 x 10⁻⁴ (****). Select pairwise p-values for Mann-Whitney tests in Tables 1–3. D.-E., quantitation of differences between near-euploid tumor cells and 2n wild-type AT2 cells. Circles, cell volume. Diamonds, nuclear volume. Crosses, N:C ratio. D., Corresponding fold change (tumor-to-normal) for cell sizes (top); standard deviation (bottom). Error bars ≤ 2.7 x 10⁻³ omitted for clarity Magenta, EGFR+, Blue, KRAS+. E., percentages of cells which fall within the regression fits to histograms for tumor cells, but outside the Gaussian fit to 2n AT2 cells; showing cell volume, nucleus volume and N:C ratio; colored by stage. Error bars, SE. Lines connect points to show trends. F., variance of cell size parameters as a function of stage. Dotted lines and solid markers, ploidy most significantly affects the variance of cell and nuclear volume. Related to Table 4 and S1 Data gives detail.

https://doi.org/10.1371/journal.pone.0274091.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. P-value results of pairwise Mann-Whitney testing select conditions shown in Fig 2C, cell volume of near-euploids.

https://doi.org/10.1371/journal.pone.0274091.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. P-value results of pairwise Mann-Whitney testing select conditions for nuclear volume of near-euploids.

https://doi.org/10.1371/journal.pone.0274091.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. P-value results of pairwise Mann-Whitney testing select conditions for N:C ratios of near-euploids.

https://doi.org/10.1371/journal.pone.0274091.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Percentages of abnormal cell volumes, nuclear volumes, and N:C ratios vs. stage, genotype, and DNA content.

https://doi.org/10.1371/journal.pone.0274091.t004

Next we asked whether the level of size dysregulation of near-euploid LA cells varied with cancer genotype or disease progression. Therefore we included in our dataset 10 patients pre-identified with either EGFR+ and 11 with KRAS+ genotypes, with 3–4 patients staged 1–3. Collapsing all stages, there is little difference between cell or nuclear volume distributions for EGFR+ and KRAS+ patients (S6 and S7 Figs). However, separating data by stage and genotype reveals subtle differences in the prevalence and trajectories of size dysregulation (Fig 2B and 2C, S9 and S10 Figs). Although we expected cell size dysregulation would worsen with stage, the independent EGFR+ (n_near-euploid = 1411, n_total = 2194) and KRAS+ (n_near-euploid = 1260, n_total = 1939) datasets both displayed the highest fraction of cell and nucleus size dysregulation at stage 2 (Fig 2C–2E, S9 and S10 Figs). Somewhat convincingly, both genotypes track together in terms of their trajectories of size dysregulation with stage (Fig 2B–2E). A significant difference between two genotypes occur at stage 1, in which KRAS+ tumor cells displayed more normal nucleocytoplasmic ratios (Fig 2C and 2D), perhaps due to slightly larger cell size and smaller nuclei compared to EGFR+ cells (Fig 2D).

Importantly, size dysregulation is not caused by a small subpopulation of cells in these LAs. Considering near-euploid cancer cells, we calculated the percent of cells whose cell volumes and nucleus volumes extended outside of the first normal AT2 Gaussian (Fig 2E, Table 4 and Methods). We found that 54.5–77.2% of near-euploid cells fall outside the range of normal AT2 cells, suggesting that size dysregulation is highly prevalent in both tumor genotypes. The extent of dysregulation is most dramatic stage 2, in which 95% of all nuclear volumes are abnormal and more than 77% of cell volumes, irrespective of ploidy (Fig 2D, Table 4). The percentage of cells with abnormal cell or nucleus sizes is also largely independent of cell ploidy (Table 4). N:C ratio abnormalities are more modest for near-euploids, although they do increase by stage 3 of LA progression (Fig 2C and 2D, S11 Fig). In EGFR+ the N:C ratio dysregulation occurs earlier than in KRAS+, reaching a plateau at stage 1 (Fig 2D). The increase in N:C ratio observed at stage 3 appears consistent with slight reductions in cell volume but considered across all cells some scaling is evident (S11 Fig). Altogether our data support partial nuclear scaling with cell volume in cancer, and broader general shifts upward in terms of nucleus size and cell size relative to the wild type.

Cell and nucleus size dysregulation is not simply explained by altered ploidy, and the extent of dysregulation differs between patients

Polyploidy is known to have a general effect on cell volume, and is a feature of aggressive cancers that is proposed to affect drug resistance and recurrence [18, 59–62]. Therefore, we asked to what extent cell and nucleus size dysregulation could be explained by changes in genome content, and whether individual patients showed distinct scaling relationships. For this analysis we plotted the volume of each tumor cell against its ploidy for each patient specimen (S12 and S13 Figs), compared to a prediction line that represents a simple multiplication of the mean DNA and volume of a normal AT2 cell. Patient samples that showed a significant number of abnormal cells above the AT2 prediction for polyploidization were classified as “supraproportional”, meaning higher cell volume enlargement than expected for the ploidy (Fig 3A and 3B, S12 and S13 Figs, see Supplemental Methods for our decision rule). Patient samples containing many cells below the line were classified as “subproportional”, or having smaller cell volumes than expected for polyploid. A majority(52.4%) of patient samples showed nonrandom enrichment of, and a majority of supraproportional cells (Fig 3B and 3C, S1 Data, tab E). Supraproportional cell volume behavior is highest at stage 2 for both genotypes (50.6% of EGFR+ cells, 49.9% of KRAS+ cells, Fig 3C, pie charts) consistent with their being the highest amount of size dysregulation (Fig 2B and 2C). By stage 3 there is increased prevalence of both “proportional” cells–cells with cell volume:ploidy that acts like simple multiplication of AT2 cell volume:ploidy—and “Near-WT” cells, near-euploids classified with cell volume within μ_WT ± σ = 582 ± 127 ratio (Fig 3A, grey box and Fig 3C, bottom). KRAS+ cells become more subproportional in stage 3, suggesting elevated ploidy without proportionate cell size increases (Fig 3B, top). Examples of supra- and subproportionality are visible in IF (Fig 3D). From these observations we conclude that size dysregulation is common in patients, and that its dependence on ploidy is patient-specific and not predominant.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Ploidy alone cannot explain a majority of cell and nuclear size dysregulation.

63X images shown. A., schematic: subgroups of scaling relationships between DNA content and cell volume: supra: cell volumes larger than expected for ploidy; sub: DNA content higher than expected for cell volume. Colors, individual patient data. Grey box, near-WT, with AT2 volume μ ± σ and near-euploid (n = 1.6–4.4). B., distributions of cell ploidy and volume for individual cells; colored to distinguigh patients. Solid black line indicates simple multiplication of the estimated cell volume and ploidy of a normal AT2 cell, or the equivalent slope; showing +/- 1 S.D., dotted lines. For example, the solid passes through the point (2 x 582 fL, 2 x 2n), where 582 is the mode of the first gaussian fit to the AT2 cell volume distribution (See S6 Fig). The three plots each represent a binning of patients from all stages and both genotypes, but with either no preference, Proportional, or a departure from this line, Subproportional, or Supraproportional. C., top, the fraction of cells classified within each category of abnormal (EGFR+, N = 10, n = 2195, KRAS+, N = 11, n = 1939, S1 Data, tab E). More cells in stage 2 display the supraproportional phenotype, and subproportional cells are enriched in stage 3 KRAS+ cancer. C., bottom, Near-WT, cells with cell volumes close to WT (455–709 fL) that are also near-euploid (1.6–4.4n) are rare in LA, but are depleted in stage 2. Magenta, EGFR+, Blue, KRAS +. D., images of high-quality KRAS+ cell models for each representative category, right, with corresponding 2D images; N, typical AT2 controls. 1, subproportional, 2 proportional, 3 subproportional. White arrows, cells corresponding to models seen in 3D renderings. Scale bar, 20 μm.

https://doi.org/10.1371/journal.pone.0274091.g003

Cell-size dysregulation is strongly correlated with loss of tissue organization and local tumor cell density

We wondered whether tissue organization within LAs might relate to the phenotype of cell size dysregulation. Gut-like morphologies–sheets of epithelial cells–are not normally found in lung alveoli, but are commonly seen in LAs [42]. Within our three-dimensional image reconstructions of LAs we identified three general categories of tumor tissue organization (Fig 4). A gut-like sheet pattern of cells we refer to “sheet growth” and is present in 61.0% of our image stacks (45/75, Fig 4, left, S1, S4 and S5 Videos, and S1 Data, tab H.). A second pattern of more “solid growth” contained highly packed cells that took on either a cuboidal appearance or one of spheroid tangled cells with elongated tails (46.7%, 35/75 image stacks, Fig 4, middle, and S2 Video). A third pattern of “sparse growth” often neighbored the other two patterns and tended to contain a lower density of cells (8.0%, 6/75 image stacks, Fig 4, right, and S3 Video). Quantification of image stacks from both genotypes predominantly containing each growth type at stage 2, shows that the extent of cell volume dysregulation correlates with tissue pattern type (Fig 4B). Both the mean and breadth of cell volume distributions tracks with tissue organization in the tumor (μ_Sheet = 1430 ± 850, μ_Solid = 1963 ± 1083, μ_Sparse = 3439 ± 3480, errors SD, Fig 4B and S1 Data tab G.). Cells with sparse organization tend to achieve largest sizes. Additionally, smaller cells tend to be more elongated, consistent with patterns of sheet like growth (S15 Fig). These data suggest that cell size dysregulation may be linked to loss of local tumor organization and cell density.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Tumor tissue organization correlates to the extent of cell size dysregulation.

A, Schematic: sheet, solid, and sparse patterns of tissue organization describe most lung tumor data. Top: representative IF signal for three types, immunostained for CK7 (red) and Lamin A+C (green), observed in both EGFR+ and KRAS+ genotypes, EGFR+ patients shown. middle: Renderings of 2D segmentations; nuclei (gray) and cell surfaces (rainbow), Scale bars, 20 μm. bottom: rendering of 3D segmentations, with stromal controls shown as bubble depictions. White box shows perspective scale, 2 μm per small tick mark. B., cell volume distributions taken from segmented image stacks similar in morphology to the canonical ones shown, exclusively from stage 2 LA and from both genotypes. Sample sizes N patients and n cells, parentheses show respective (EGFR, KRAS) samples: N_Sheet = 4 (1,3), n_sheet = 155 (63,92), N_Solid = 4 (2,2), n_Solid = 438 (275,163), N_Sparse = 2 (1,1), n_Sparse = 73 (22,51).

https://doi.org/10.1371/journal.pone.0274091.g004

Human lung tumor cells of near-euploid DNA content can be purified on the basis of cell size

To confirm cell size dysregulation occurs in near-euploid LA and rule out artefacts related to imaging or sample preparation, we decided to isolate cells from resected tumors, select for cells of lower, near-euploid DNA amounts, and sort them by size using flow cytometry. Building on a general framework for separating human cells correlating their dimensions to light scattering profiles [63], we developed a FACS-based strategy for separating near-euploid human LA epithelial cells on the basis of their sizes.

This approach combines three protocols developed previously: a gentle digestion cocktail to preserve CD45 receptors for removing blood cells [64, 65], fixing cells to preserve their shape in the flow cytometer, and quenching fixative with glycine to prevent cell clustering [66]. Our cell suspensions were CD45- after depleting immune cells. Importantly we determined that for epithelial cells from LAs, side-scattering width (SSC-W) parameter was proportional to cell cross-sectional area (A_Cell = (-159 ± 44) + (4.35 ± 0.48)SSC-W μm², N = 12, S16 Fig) [63].

To enrich cells of interest we gated for live and dead cells using Zombie Red, epithelial cells using EPCAM, and lower ploidies using the low-DNA DRAQ5- gate (<1.5x mode, Fig 5A). We further separated cells into size bins using SSC-H vs. SSC-W gates as shown (Fig 5B, SSC-W = 64–79, 84–99, 101–113). Reassuringly, cells were in clearly distinct bins of SSC-W but had normal DRAQ5 content (Fig 5C). After sorting, the bins showed clear separation by cell area (Fig 5D). As expected, normal lung tissue displays only modest shifts in size distribution and a small yield of cells in the largest cell-size gate (Fig 5D and S17 Fig). SSC-A vs FSC-A plots show that FSC-A does trend upward with SSC-W gates G1-G3 as expected for objects of increasing cross-section (S18 Fig shows a traditional). These results are consistent with those from whole mount tumor imaging, and further demonstrate the high prevalence of epithelial cell size dysregulation in LA, independent of high ploidy.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Near-euploid tumor cells can be purified and sorted by cell size using flow cytometry.

A. 9.43 x 10⁵ total events shown for a dissociated tumor from patient 5296 of unspecified genotype and grade. Signal gates to purify epithelial cells that were alive when fixed, and with near-euploid DNA amounts are shown left-to-right overlaid on fluorescence histograms for the parent digest, top row, and CD45- MACS flow-through, bottom row. The DNA gate is 1.5x the mode shown for CD45-. B., The SSC-H vs SSC-W plot is gated as shown to separate Zombie Red-/EPCAM+/DNA- cells selected from A. C., Contour plot of the SSC-W as a function of DNA for Zombie Red-/EPCAM+ cells, green, as compared with DNA-size-sorted cells as shown. Size-sorted tumor cells have SSC-W that is essentially independent of DNA amount. D., Representative histograms of cell areas measured from brighfield images of cells on a hemacytometer as shown in E., for normal lung tissue, left, and a LA tumor, right. Top to bottom, each row shows the parent, the MACS CD45- flow-through, and the results from the three size gates. Percentages, relative cell yields for gates G1-G4. G4 always contained low yield aggregates. E., brightfield images of cells sorted by gates G1-G3, and with mean cell areas close to those shown in 5D. Scale bar, 20 μm.

https://doi.org/10.1371/journal.pone.0274091.g005

Cell-size dysregulation is AT2-like, and correlated with loss of a novel network of cellular processes

In another attempt to explain cell size dysregulation in LA, and given the observation that sheet-growth cells are smaller and retain size regulation, we considered whether cell size dysregulation in is in any way linked with preexisting pathways for differentiation. AT2 cells differentiate into AT1 cells through a process of flattening coupled to a loss of expression of the AT2-specific marker prosurfactant protein C (proSPC, [45]). As AT1 volumes are unknown, we considered whether cell size dysregulation results from AT1-like differentiation, sans flattening. However, RAGE, a marker specific for AT1 cells in normal tissue showed no signal in LA cells (Fig 6A). In addition, we observed significant flattening of LA cells only rarely (S19 Fig). Thus, our data are not consistent with a model of AT1-like differentiation leading to cell size dysregulation in LA.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. AT2-like cells display cellular processes whose organization is altered in lung adenocarcinoma.

A., 25X images at 0.3779 μm resolution showing a single plane display RAGE, (green) is expressed highly in AT1 cells in normal tissue, top, but we could not detect expression in either genotype of tumor cells bottom. KRAS+ shown. Cells immunostained for CK7 (red) and DNA (cyan). B., 63X images shown. Left three columns, raw data shown in a maximum-intensity projection (MIP) 5 μm thick illustrates the behavior of the AT2 process network in normal distant controls, top, and tumor lung, bottom. Right column, a snap from the Imaris workstation of a MIP over an entire processed image stack shows the overall structure of the process network. Each row shows a different specimen (see Table 5). proSPC (green), CK7 (red). Bottom row, transition regions are observed in both genotypes of LA and contain presumptive diseased AT2-like cells that neighbor enlarged CK7-expressing cells. EGFR+ tumors shown (also S4 and S5 Videos).

https://doi.org/10.1371/journal.pone.0274091.g006

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Patient identifiers corresponding to data used to create each image shown and select figures.

https://doi.org/10.1371/journal.pone.0274091.t005

Next we considered the alternative possibility that LA cells express AT2-specific markers and show morphological changes characteristic of AT2 cells, as suggested by results in mice [44]. Fortuitously, IF revealed a previously unreported feature of human AT2 cell anatomy, along with a corresponding pathology in LA cells (Fig 6B). We refer to these structures as “processes” or “process networks”. Processes are robust to antibody titration and display incomplete colocalization with E-cadherin and CK7 (S20 and S21 Figs). We supposed that if processes are a cross-reaction with extracellular matrix (ECM), they should disappear during a mechanically gentle collagenase/dispase digestion over several hours. However, observations revealed they are clearly not elements of ECM (S6–S12 Videos). During digestion, processes remain wrapped in shafts of intracellular cytokeratin (S11 Video), can remain connected between AT2 neighbors (S8 and S9 Videos), and can even be freed from ECM entirely (S7 and S12 Videos). Although exceedingly fragile, processes loosely follow the paths of CK7 bundles and show a level of complexity not typical of cross-reaction artefacts (S21 Fig, S4A and S4B Video). Amino acids 1–100 of proSPC that were the antigen for our antibody do overlap with the 35-residue region that makes up only 10% of the mass of the secreted form, yet we observe no significant reaction with the inner alveolar walls [48]. We hypothesize that processes are components of the cytoplasm of human AT2 cells. We do caution that a second antibody frequently used in mouse studies, raised against residues 1–32 of human proSPC was less effective in our hands and was not able to detect processes like those shown (S20 Fig and Table 6). Neither antibody we tested detected processes in normal or LA mouse samples. Thus, if mice possess this trait, an antibody raised against the mouse propeptide 1–100, with identity of 86% with the human sequence, may be required to reveal these structures (S20 Fig). In human LA, process-like staining is visible in sheet and solid growth patterns (Figs 1C and 6B, S22 Fig). Sparse CK7-positive cells are usually visible in proximity to sheet and solid regions, where they lack process-like staining (S22 Fig). Conversely, collagenase-resistant clusters of LA cells retain process-like staining at their interior junctions (S13 Video). These results support a link between cell size dysregulation and alterations in surfactant physiology, and strongly suggest human LA cells are AT2-like.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Primary and secondary antibodies used in this study.

https://doi.org/10.1371/journal.pone.0274091.t006

Significance of dysregulated cell and nucleus sizes to cell and tumor biology

The high prevalence of dysregulated cell sizes in LA suggests that these cells are contributing to tumorigenesis. However, an alternative hypothesis is that these enlarged cells are no longer functional or proliferative and perhaps senescent. To test this, we measured the average nucleus size of proliferating normal and LA cells across genotypes and stages. We found that the average nucleus size of proliferating cells was dramatically increased in LAs (Normal, 7.64 ± 0.04 μm, EGFR, 11.4 ± 0.12 μm, KRAS, 11.1 ± 0.12 μm error SEM, p = 2.5 x 10⁻⁶⁹), suggesting that these enlarged cells are not senescent but are contributing to tumor growth (Fig 7A, S23 and S24 Figs).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Enlarged EGFR+ or KRAS+ lung adenocarcinoma cells can proliferate, possess similar protein density to stroma, and display organelle size scaling.

A. Nucleus sizes from proliferating EGFR+ and KRAS+ tumor cells identified by Ki-67 staining, relative to non-proliferative normal AT2 cells (N_EGFR = 8, N_KRAS = 8, n = 435). B., Zombie Red (red) is a broadly amine-reactive dye that in these FFPE samples reacts equally with small stromal cells, S, as it does with enlarged cancer cells, T, which is inconsistent with cells swelling via water uptake (N = 2). DNA stained with TO-PRO-3 (cyan). Right, Zombie Red intensity of KRAS+ tumors, a proxy for protein density in the cell, does not show the downward trend with cell area expected for water-diluted cells (N = 2, n = 20). C. Allometry of nuclear volume as a function of cell volume for LA (N = 21, n = 2082). Red line, double exponential fit to data. D., Fibrillarin (green) marks nucleoli comparing tumor cells of varying size in a KRAS+ stage 3 patient’s tumor (ID# 4855). CK7 (red), TO-PRO-3 (cyan). E., Total fibrillarin content in nuclei scales with segmented nuclear volume as expected for productive nucleoli (related to D., N = 1, n = 212). F., SON (green) marks nuclear speckles (also KRAS+ patient 4855). CK7 (red), TO-PRO-3 (cyan). G., Quantification of total SON content shows speckles also scale with segmented nuclear volume as they do in living cells (related to G., N = 1, n = 125).

https://doi.org/10.1371/journal.pone.0274091.g007

A recent study proposed that water uptake is a consequence of aneuploidy, based on stiffness measurements of culture cells and plots of cell-size vs. ploidy for the NCI-60 panel [30, 31, 67]. To simplify our analysis of swelling, we considered two extreme cases. In the first case, water uptake fully explains cell volume differences predicting a linear negative correlation between protein density in cytoplasm and volume (Fig 7B, Case 1). In the second model, protein density is constant with cell size, implying no dependence (Fig 7B, Case 2). To test these models, we examined both intensities of Zombie Red, which reacts with all lysines, and CK7, a cytoskeletal protein that scales expression with cell volume [22]. The Zombie Red analysis shows that relative intensities in cytoplasm are not dissimilar from smaller, immediately-adjacent stromal cells (Fig 7B, compare S and T), but does not necessarily support either model. CK7 staining also supported neither swelling nor independence, but did suggest the very largest cells rarely possess the upper 50% of possible protein density (8.3% of cells >2 pL, 18.8% of cells ≤ 2 pL, S25 Fig). Rounded bodies and nuclei are another defining characteristic of turgid cells in hypotonic medium [68]. Such morphologies do exist within our samples, but are not conspicuous (e.g., a handful of nuclei in Fig 4A, center and right are representative). Consistent with this interpretation, cell model sphericity was 0.47 ± 7.1 x 10⁻⁴, with sphericities of ≥0.6 present in only 0.6% of all measurements (n = 4133, S1 Data, tab B). Similarly, lamin A+C staining of nuclei showed highly irregular surfaces (final portions of S2 Video, Fig 4 and S26 Fig). Nuclear models show a minority of oblate (15.0%) or prolate (26.0%), ellipticities less than 0.3. Overall, our data supports a view in which a minority of LA cells swell, but in a manner that is not proportional to cell size.

For enlarged cells to function and proliferate, they would need to expand their organelles and output of pathways for RNA production and ribosome biogenesis. We first noticed that enlarged cells properly maintained nucleus size scaling (Fig 7C). To determine whether nucleolus size and rRNA processing might be upregulated, we stained LA samples for fibrillarin. Indeed, we found a clear size scaling trend in which nucleolus size was increased in epithelial cells with enlarged nuclei (Fig 7D and 7E). To determine whether splicing and RNA production might also be upregulated in enlarged cells we stained for SON, a splicing marker that localizes to nuclear speckles [69, 70]. These nuclear bodies also showed a clear size scaling trend with nucleus size (Fig 7F and 7G). Together these data indicate that enlarged LA cells minimally maintain their protein and organelle homeostasis, scaling it as a function of cell size. These results contradict the notion that these enlarged cells are only swollen or senescent and hints at their higher functional output.

Genomes of oversized lung adenocarcinoma cells are physically expanded

Given the correlations observed between cell and nuclear volume, and findings that changes in chromatin organization can be observed in cancer, we asked whether DNA staining displayed any special patterns in LA nuclei. We observe five common classes of DNA staining, and mixtures of these that describe most DNA data in LA (Fig 8A). Normal AT2 cells have mostly even DNA-staining profiles, with little granularity, and are so consistent that neoplasms are usually conspicuous. In LA, the most common pattern consists of a granular appearance suggestive of local heterochromatic regions. Cytoplasmic inclusions are ringed by concentrations of DNA at their boundary. Lamin-associated DNA is concentrated at the nuclear periphery and sometimes at the periphery of inclusions. DNA voids consist of large volumes that stain poorly for DNA, but are otherwise not significantly bordered. “Wagon-wheel” arrangements are reminiscent of nucleoli, but may not contain them, as suggested by weak DNA staining at their central “hub” regions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Diversity of nuclear organization and a revised 2D perspective of human lung adenocarcionoma.

A., typical patterns of DNA staining observed in both EGFR+ and KRAS+ LA. n, normal AT2 cells with proSPC (yellow), CK7 (red), DNA (cyan). gr, granular DNA. cy, cytoplasmic inclusions, with fibrillarin (green). lad, lamin associated DNA, with Lamin A+C (magenta). v, DNA voids. w, “wagon-wheel” DNA. Idealized representations of nuclear structures shown for clarity. See Table 5 for associated patient ID information. B., Visual depiction of the relative equatorial cross-sectional area in μm² of typical LA cells, if they were spheres, relative to the nucleus (cyan). green, AT2, cyan, DNA, black, DNA voids, magenta, nucleoli/speckles.

https://doi.org/10.1371/journal.pone.0274091.g008

We took particular note of Lamin association and thus we asked whether nuclear size dysregulation is correlated in any way with Lamin A+C wrinkling. We observe high Lamin A+C expression in all our patient samples, as has been observed previously for the LA subtype [71]. Lamin A+C is such a strong marker that it enabled segmentation of most of our dataset with the membrane detection algorithm of Imaris^™. This algorithm creates objects that are significantly smaller than the nuclear volume, but contains very accurate shape information about the sphericity, i.e., wrinkling, of the nuclear envelope (S26 and S27 Figs). We reject any significant relationship between lamin wrinkling and nuclear volume, and by proxy cell volume (S28 Fig). However, Lamin A+C data support a similar trend as nuclear volume data with disease stage as measured by DNA thresholding, supporting that method (Fig 2D).

The presence of DNA-poor voids led us to ask whether genomes could be diluted in larger cells. To test this, we plotted the cell volume as a function of the ratio of ploidy:nuclear volume, where the latter tracks the amount of genome dilution (S14B Fig), showing that DNA in large LA cells >2 pL is diluted within the nucleus, consistent with voids.

Taken together with the above, and with the observations of lamin-associated DNA, we wondered if DNA movement to the envelope, sometimes associated with gene regulation, may be a hallmark of oversized cells [72]. To answer this question, we used the distance transform to create eroded shells of nuclei for four patients, found the amount of DNA within 1.5 μm of the nuclear surface, and then calculated the amount of nuclear-envelope enrichment of DNA as a function of cell volume. However, there is no trend with cell size, but shell-packing of DNA staining may be a signature of individual patients (S29 Fig).

Fig 8B summarizes how this study has revised our expectations about the 2-dimensional length scales of cell bodies, nuclei, nucleoli, and nuclear voids of AT2 or LA cells of different sizes, if cells and nuclei were actually perfect spheres.

Conservation of size dysregulation in a mouse model lung cancer system

We were curious whether the phenomenon of cell and nucleus size regulation might be specific to human LA or might also be found in a mouse model of lung cancer. We segmented normal nuclei and also cancer nuclei from proSPC-staining cells in a KrasLSL-G12D/+; p53fl/fl; Rosa26LSL-eYFP/LSL-eYFP mouse. Our analysis showed that although mouse AT2 cells are smaller than human AT2 cells, both show loss of size homeostasis in cancer (Fig 9A). Mean nuclear volume for the mouse was 117.4 ± 21.36 (n = 58), 1.72-fold smaller than the first Gaussian of human AT2s. For tumor cells, which were filled with proSPC at all sizes, the nuclear volume was 254.7 ± 358.5 (n = 207), or 2.17-fold smaller than the overall human LA mean. Importantly, the mouse coefficient of variation for LA nuclear volume was 1.41, somewhat higher than the human average for all LA cells of 0.68. Altogether, the human model possesses a greater range of nuclear volumes and is more relevant to human disease, whereas the prevalence of volume variation in the mouse model provide opportunities for genetic characterization of pathways of size variation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Epithelial cell-size dysregulation is also found in mouse tumors and may proceed with loss of tissue organization and upregulation of transcription.

A., KrasLSL-G12D/+; p53fl/fl; Rosa26LSL-eYFP/LSL-eYFP mouse lung tissue shows nuclear size dysregulation, despite volumes of ~50% that of humans. B., Model 1: Oncogenesis is characterized by lesions, loss of flattening to an AT1 form and thus shortening of processes. At stage I, a contact-dependent mechanism preserves some cell size regulation. Stage II involves more mixed stroma/tumor regions and disorganization, leading to sparse cells and increased cell size dysregulation. Late stage (III-IV) cancers are dominated by adaptation towards maximal cell division rate, evolution of immune resistance, and also loss of cytokeratin staining. Smaller cells, and those of the more subproportional volume:ploidy relationship in KRAS+ cancers (Fig 3) may become enriched in late stages.

https://doi.org/10.1371/journal.pone.0274091.g009

Patient samples, genotyping, formaldehyde fixation and paraffin embedding

Formaldehyde-fixed paraffin-embedded (FFPE) samples were kindly provided by the University of Pennsylvania Tumor Tissue and Specimen BioBank (TTAB). Samples included ten blocks from untreated patients genotyped as EGFR-positive/KRAS-negative and eleven blocks from patients genotyped as KRAS-positive/EGFR-negative. Staging was determined by standard measurements of tumor diameter and interpretation of H&E by a pathologist and were as indicated in Table 7.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Stage, genotype, and pathological classification of lung adenocarcinoma patients.

https://doi.org/10.1371/journal.pone.0274091.t007

Fixation was performed in neutral buffered 10% formalin for a minimum of six hours, with all steps including stepwise dehydration, clearing with xylenes and paraffin-embedding performed on a commercial tissue processor. Processing was performed by the Hospital of the University of Pennsylvania Department of Pathology and Laboratory Medicine Division of Anatomic Pathology.

Normal distant control tissue was identified and marked by Dr. Feldman by inspection of H&E clinical slides. FFPE blocks were then subdivided, such that distant control regions were used for imaging normal AT2 cells.

Dewaxing and sectioning

Up to 12 FFPE blocks, including 6 distant controls and 6 tumor specimens, were dewaxed for 24–72 hours in two six-well aluminum Reacti-blocks (Fisher, TS18811) immersed in 3 L of xylenes (Fisher Chemical, X3S-4). Samples were next washed once with 50% ethanol and then twice in anhydrous ethanol for storage at -20°C. Storage was for up to two months for principal IF in which channel 1 was proSPC, Lamin AC, Fibrillarin, or SON, up to six months for Ki-67 proliferation testing. S30 Fig shows that IF was stable over the period of principal imaging for both normal controls patient 4844, with 25X images shown in Fig 7B and S6K Fig representative of sample response after 6 months storage. For whole-mount sectioning, specimens were rehydrated to 50% ethanol for 5 minutes with nutation followed by one wash with water and one wash with PBS. Human specimens were sectioned to 150–200 μm on a Leica Vibratome (VT1000S) in PBS, mounted with superglue but otherwise with no solid or gelatinous infusion such as agarose, with maximum amplitude settings and frequency of from 50–100% depending on tissue softness. i.e., 100% normal lung distant controls required the highest setting. After sectioning, specimens were again dehydrated for storage at -20°C.

Antigen recovery, immunofluorescence, clearing and mounting

All staged and genotyped tumor FFPE blocks were kindly provided by the University of Pennsylvania Abramson Cancer Center Tumor and Biospecimen BioBank (TTAB). FFPE sections stored in ethanol at -20°C were rehydrated as described. Antigen recovery upon 150 μm sections was performed in loosely capped 5 mL reaction vials using a Decloaking Chamber (BioCare Medical) at 110° for 37 minutes using a 10X TE buffer at pH 9 at a 1:100 dilution in dI water (Dako Target Retrieval). The sample was allowed to cool to room temperature and then washed 3 times with 0.5X SSC buffer (Corning). Samples were transferred to bleaching buffer (0.5X SSC, 2% hydrogen peroxide, 5% formamide) and bleached for at least 4 hours under a 130 V 100 W bulb at a distance of 20 cm at 4°C. Samples were returned to RT, then washed three times in quick succession and again every hour for four hours with PBT (PBS plus 0.1% triton and 2 mg/mL BSA). Blocking was performed for 2 hours at RT in blocking reagent with nutation. Blocking reagent was a 9:1 mixture of PBS containing 0.1% triton X-100 and no BSA and goat serum (abcam 7481). We stress that any BSA in blocking reagent specifically interfered with IF in our hands. Primary antibodies at the concentrations listed in Table 6 were applied in blocking reagent overnight in 2 mL glass reaction vials with nutation at 4°C. On the second day, samples were washed with PBT at room temperature on a rotator over a 7-hour period with the wash changed at least three times at intervals of at least 1 hour. Secondary antibodies as shown in Table 6 were applied overnight in the dark at 4°C in blocking reagent with nutation. On day three, the samples were again washed on a rotator for 7 hours in PBT. TO-PRO-3 at a 1:500 dilution in PBS was applied in one wash for 20 minutes on a rotator and then refreshed for 10 minutes such that a faded blue color remained in the solution, which we took as indication that DNA binding sites were saturated. To prevent washout of TO-PRO-3 stain that we have observed in mixtures of water/methanol, samples were immediately dehydrated to 100% methanol for 30 minutes with nutation, then washed again once with methanol and once with anhydrous methanol at intervals of 30 minutes. To clear tissue, samples were placed in 1:2 Benzyl Alcohol:Benzyl Benzoate (BABB) on a rotator in the dark for 10 minutes. Next the initial BABB was replaced, and samples were cleared overnight in a dark drawer. The next day the BABB was exchanged and the samples were cleared for another 24 hours. BABB was refreshed a final time at mounting. For mounting, samples were subdivided 1–3 times with a razor to lay flat, then placed in custom laser-cut chambers with 3.75 x 19 x 19 mm windows filled with BABB. Samples were pressed under a second 18 x 18 mm coverslip to lay flat before the chambers were filled to level. 40 μL of BABB was removed to leave a small bubble and ensure no overflow before the chamber was finally sealed with nail polish.

Laser-scanning confocal imaging

Principal imaging relevant to all cell volume segmentation and high-resolution figures was performed at room temperature on a Zeiss LSM-880 with a Plan-apochromat 63X/1.4 oil DIC M27 objective with 0.19 mm free working distance, 0.17 mm coverglass thickness, 25 mm FOV, and 45.06 parfocal length. Scanning settings were 10 (fastest setting), with 900 x 900 pixels and 0.15 μm³ voxels, employing a 1X digital zoom. Laser lines were 488, 561, and 633 nm, employing transmittances of 5, 1, and 1% respectively, and each used a gain setting of 600.

Efforts were made to minimize sampling bias during imaging, but only regions staining brightly with either proSPC, for normal AT2 cells or mouse tumor cells, or CK7 (bonafide tumor cells) were imaged. To capture representative morphological patterns, a number of 25 X images were snapped to capture all the overall morphological variation present in a sample. This initial impression was used as a guide to gather from 1–4 scans at 63X until all the bright-staining representative morphologies available were captured. Any 63X image stack was constrained to include a minimum of 20 stromal nuclei controls for DNA amount. Therefore, cells with phenotypes which stain weakly with either proSPC or CK7 are not included in our data.

25X images were collected similarly to 63X images, except that the voxel size was 0.3779 μm³ and z-step of either 2 (Ki-67) or 0.5 μm (others).

Attenuation correction and background subtraction of raw data

Attenuation is loss of signal as light scatters nonspecifically by interacting with matter as it passes through the object of interest. As a result, the signal from the fluorophores diminished as a function of z-depth. Data for channel 1 (488 nm) did not display any attenuation in our hands. Data for channels 2 and 3 displayed attenuation of differing degrees. To estimate the attenuation, images were opened in FIJI, and the mean intensity over at least 100 pixels was measured in a uniformly staining square region of a typical brightest object (e.g., nuclei) present near both the top and bottom of each image stack. For example, if a normal multilobed neutrophil nuclei was observed at the top of the stack that was the brightest observed, then a similar neutrophil was measured near the bottom of the stack that was also the brightest observed. If normal cells were not available at the top of the stack, then similar cancer cell features were selected, such as DNA concentrated near the outer edges of nuclei. Attenuation measured for channel 2 (CK7) was typically less than 20%, and attenuation for channel 3 was consistently about 2-fold over approximately 350 z-steps. To correct attenuation, mean intensities for top and bottom were used in combination with the Attenuation Correction MATLAB plugin for Bitplane Imaris 9.2.1. Background subtraction was performed in Imaris 9.2.1 using the Background Subtraction function in the Thresholding submenu of the Image Processing menu, with all three channels checked and the default parameter of 33.7 um filter width applied.

Segmentation of normal AT2 cells using Bitplane Imaris^™ software

Confocal IF image stacks of distant control samples stained with anti-proSPC and anti-CK7 were selected for segmentation if they contained alveolar arrangements in the CK7 channel. Cells were classified as non-neoplastic if they fulfilled the criteria of both presenting proSPC in a dense punctate arrangement outlining a cell body, extending into a network of long processes, as well as possessing signature DNA staining patterns lacking any inclusion structures, large voids, or granules (Fig 7H). 3D segmentation of cell bodies and nuclei was performed using standard manually guided procedures in the wizard of Bitplane Imaris^™ 9.2.1, combined with some channel masking and erosion steps, and is detailed in the walkthrough in S4 Fig.

First, a surface object was built with a coarse level of 1 μm smoothing, with an isovolume threshold cutoff of from 7–12 applied to the proSPC staining such that the general form of the cell bodies and their nearby neighboring space was captured. Filters were applied to remove thresholded fragments of AT2 processes and cells at the borders of the imaging area. The resulting masked DNA channel highlighted inappropriately included neighbors and was used as a guide to erode the coarse SPC surface using the distance transformation plugin of Imaris^™. The initial coarse cell body model was then eroded in 3D by from 0.7–1 μm, as minimally as possible, but as appropriate to minimize inappropriately included neighbor nuclei. This erosion of the coarse proSPC surface was then applied to mask the DNA channel, removing most of the inappropriately included neighbor nuclei.

The resulting masked DNA channel was used to create a near-accurate surface from both AT2 and some clipped neighbor DNA. To accomplish this a cell object was created with settings of 0.3 μm smoothing, threshold of 7, and splitting activated with 2 μm seed points and all seed points accepted. These surfaces were exported to a surface object and the resulting surface containing over-split nuclei and a few nearest-neighbor clippings were manually rejoined and shaped, using the editor with liberal use of the surface scissor tool, or deleted as appropriate, always with 3D visualization of the data as a guide. After manual editing, small clippings with near-zero volume or area resulted which were filtered and deleted. The resulting segmentable AT2 nuclear surfaces were then used as a final accurate mask for the DNA channel.

For final models of the cell body, the previously described erosion surface was used to mask the proSPC channel. This resulting masked channel for proSPC was added to the DNA channel that was masked previously by using the segmented AT2 nuclei. This channel addition was performed by using the channel arithmetics plugin. This final segmentation channel, which provides a full signal from the entire AT2 cell body, was used to create the 3D segmentations of the AT2 cell bodies. For this model of normal AT2 cell bodies, a cell object was created with the following settings. 0.3 μm smoothing, a threshold of 7–12 chosen with manual inspection using Imaris^™ 3D view and wizard, and surface splitting with 5 μm seed points. Manual curation was performed on the cell body surfaces as described for the nuclei, with continuous reference to the unmasked SPC channel. To create fully segmented cells, a new cell object was created and both the cell body and nuclear surfaces were imported. S4 Fig summarizes this process and S1D Fig is representative of the high accuracy of this method relative to the pattern of proSPC staining.

We wish to note that AT2-like lesions or small neoplasms that present proSPC, but are sparse and border CK7+ tumor cells cannot be segmented by this method or the method described below for CK7+ cells, and are thus excluded from our analysis (S4 and S5 Videos).

Image annotation for training set generation

To train our UNet-based convolutional neural network (UNet), we settled on an empirically guided, data-driven annotation methodology. All annotations were performed guided by the CK7 channel of an autoscaled merge of red CK7 and cyan DNA. We employed a conservative, low-inference boundary drawing approach, performed only on obvious cells with clear boundaries, which in tests produced the strongest predictions (S2 Fig). To establish a high-quality training set of 40 images, Dr. Sandlin annotated 80 images and a second expert in cell biology, Dr. Good evaluated the images with a score of 0–6 based on the representativeness and accuracy of the cell boundaries drawn. The highest-scoring 40 images from the original 80 with the clearest cell boundaries and brightest CK7 signals were selected for our final training set. Further testing revealed that masking the data supplied to the network external to the boundaries and eliminating non-cell marks improved prediction quality based on cell coverage and qualitative reduction in apparent false-positive marks (S1C Fig).

Prediction of tumor cell boundaries using a 2D U-Net-based convolutional neural network

S1A Fig illustrates the architecture of the U-Net [88]. A deep convolutional neural network was trained with the aforementioned set of 40 images in order to detect cell boundaries. Training images with initial size of 900×900 were cropped to a size of 448×448 to obtain 160 images. Data augmentation was used to increase the size of this dataset in order to prevent over-fitting. This augmentation involves the application of rotation, shift, shear, zoom and flip to each annotation image. These operations increased the training set to 4960 images. This training set was normalized by dividing the 8-bit images by 255 and centralized by subtracting the mean over the z-dimension of pixel values for each image. In addition, the training set was split into 2 parts. 80% of the training set was used for training, and 20% for validation. A loss function was employed, with the form (1) where L is the binary cross-entropy. The optimizer used during training was Adam, and Dropout was used to prevent over-fitting [89, 90]. The criteria used to stop training and save the model was a cutoff for the loss function 1. of less than 0.001. UNet code is available on request from the corresponding author, and may be used with cytokeratin IF images to detect cell boundaries.

3D segmentation of tumor cells using Imaris software with UNet boundary predictions as input

S3 Fig illustrate the procedure. Cancer nuclei were manually annotated with a white mark in FIJI based on the presence of CK7 stain and recognition of abnormal DNA morphology. Cells which appeared cut off by the edge in the 2D image were not annotated. This white mark was used to approximate the positions of nuclear centroids using a custom MATLAB script to simulate an idealized spherical nuclear object with radius 5 or 10 pixels. Raw.czi files at 900 x 900 that were preprocessed as described and converted to.ims files were cropped to 896 x 896 by removing 2 pixels from each edge using the 3D crop interface in Imaris. The 896 x 896 U-Net prediction stacks and the idealized nuclei represented by white spheres were then imported into Bitplane Imaris^™ 9.3–9.5. A cell object was created with the nucleus detection scheme, employing the aforementioned idealized nuclear marks with the U-Net predicted boundaries as an input for the nucleus channel. For nuclei detection, an object radius of 1.6 μm was used for 5 pixel radius spheres and one of 2.75 μm diameter was used for 10-pixel radius spheres. A smoothing factor of 0.3 μm was used in all cases and a threshold of 10 units was employed for the idealized nuclei isosurface. For cell bodies, we used the membrane detection algorithm. According to the training provided by Bitplane^™, in the slice view, we measured the diameter across the smallest predicted cell boundary for a tumor cell and estimated an even value of 4–10 μm. Due to the high variability in the cell sizes in LA, a second fit was sometimes performed using a larger diameter of 4 um larger, although with this step we rarely observed an increase in yield of more than 1–2 successful cell models (<5%). All voids in the model with angular appearance fit by the cell module or those that touched the edge of the image were deleted, including all cells with more than an apparent ~10% of their volume cut off by the image edge. This cell object was exported as a surface object, made 50% transparent colored by ID and used to mask and duplicate the DNA channel. Using the 3D view, all the tumor cells that contained fragmented nuclei within them or extensive clipping into neighbors (~10% of a neighbor nucleus) were deleted as failed segmentations. Double nuclei were retained as correct if their boundaries were obviously fused or the overall morphology of channel 2 CK7 supported this choice, but in most cases double nuclei were deleted as failed segmentations.

A second round of manual curation involved creating a new cell object and saturating the CK7 and DNA channels (upper limits of 100 and 50% max, respectively). In combination with the clipping planes feature of Imaris^™, we were able to visualize in the xy and xz or yz planes how the cell bodies as volume representations agreed with the CK7 and DNA staining. In this way, most inaccurate segmentations were removed from the data.

A third round of manual curation involved inspection of all cells with N:C ratios observed in the top 10% and bottom 10% of the data. These suspect models were revisited in 2D and 3D and were deleted if their models were erroneous in terms of significant overlap between nuclear segmentation and cell boundary, in either the 2D slice view or in the 3D Imaris slice view, in any of the X, Y, or Z directions (S3B Fig, frame 2).

A final round of manual curation involved generating 2D line representations every 5 μm along the z-direction and observing agreement between the cell body and nuclei segmentations and the channel 2 and channel 3 data. If significant errors in 2D area were noted, cell models were revisited using the clipping plane method in 3D as described and deleted if they disagreed with the UNet model or image data.

For a minority of cells with complex shapes that were viewed as erroneous in their initial models, but nonetheless possessed strong predictions, we attempted to recover the models for volume estimation by refitting the UNet boundary in a small ROI, and then combining the surfaces using the unify tool.

3D segmentation of tumor cell nuclei using Imaris^™ software

Initial tumor cell body models were exported again as a surface object used to duplicate and mask channel 3 again. Due to the extremely crowded and irregular arrangements of tumor cells, even slight inaccuracies in cell-boundary prediction can still result in minor amounts of neighbor clipping. To eliminate the possibility of clipped neighbor DNAs affecting our nuclear segmentations, we masked these clippings with a new cell object, in a way that was similar to the procedure for AT2 nuclei. This clipping mask cell object was fit to tumor cell DNA with the following parameters: Cell body only detection, Smoothing 0.3 μm, Background subtraction 5 μm sphere diameter, cell body detection, threshold 1 unit, Split by Seed points 2 μm, with all seed points accepted. Clippings were thus negatively selected for and masked away via extensive manual curation. To accurately model nuclei, we found that our Lamin AC stain was insufficient, because the fitting algorithm takes the inner shell of the Lamina and always underestimates the size of tumor nuclei (S5 Fig, white surfaces). Instead, we chose to use the low-background far-red channel 3, strongly present in all our data. However, we found that the volume of an isosurface created in Imaris is dependent on the overall brightness of the image stack, as represented by the center-of-mass of the channel 3 histogram, as well as the threshold chosen (S5 Fig, magenta and red surfaces). To attempt to correct for this minor empirical drift, we sought a self-consistent rule by which to model the isosurface, in such a way that the surface would match the expected size and shape of nuclear envelopes, but also be of sufficient volume to encompass all the DNA data for quantification. To derive such a rule, we determined a linear relationship between the observed distance of a given isosurface based on DNA intensity, with dimensions that themselves depend on threshold, from an ideal Lamin AC surface. We now describe how we used available Lamin data to derive an empirical relationship between the threshold which we selected in Imaris to model nuclei, upon a given DNA channel intensity histogram’s center-of-mass.

Let t be the threshold selected to build an isosurface based on DNA stain, l the amount by which a lamin envelope surface is dilated to make a larger isosurface, and d be the DNA content inside of the resulting 3D surfaces. We approximate the DNA content as identical between these two isosurfaces, which are based on DNA or lamin dilation, for some corresponding values of t and l. For a given DNA channel histogram center-of-mass, I, the experimentally observed behavior of these dna-content functions, d, for small changes is approximately linear, obeying (2) (3)

I a constant for one image stack, where the β, m are the respective intercepts and slopes (see also, S5A Fig to understand the differences in sign). Restated in simpler terms, the dna content d inside each isosurface goes down when we increase t and up when we increase l.

We seek an empirical function for the dependence of the threshold we select, t, on some theoretical amount of lamin dilation (yet to be decided), which we will call L. From 2 and 3, we have rates (4) (5)

Therefore, (6)

This provides the slope of the desired function. However, we need to determine the form of t(L), such that as L increases, t decreases. The intercept β_t represents the theoretical dna amount contained in the isosurface at zero threshold and as L → ∞. This intercept is also the maximum value at which d(l) should be defined, because it is the amount of dna contained in a nucleus within a 3D reconstruction of infinite size. For L = 0 (no dilation of the lamin-based surface), we have d = β₁, which we know to underestimate the real amount of DNA (see S5A Fig, Lamin) This results in range and domain for L and d of (7) (8)

From these values we can use 2 to determine a range for our desired function t(L).

So the form of the function we are constructing should depend on L, take on maximum value , and have slope . Because m_t is negative, for clarity we use the absolute value below. This gives the function (9)

The dna content value we report for tumor cells in this work is (10) arbitrarily chosen as 90% of the theoretical maximum DNA that would be associated with zero threshold and infinite volume. S5A Fig illustrates how this choice qualitatively fits nuclei for different histogram brightnesses. The associated dilation amount, L, calculated from 3 and 10 is (11)

we calculated the appropriate thresholds associated with these values by substituting 11 Into 9, giving (12)

To determine the relationship between t90 in Eq 12 and histogram brightness, I, we chose several files of varying brightness and measured the DNA content at different thresholds and lamin dilation amounts. Dilation was accomplished using Bitplane Imaris with the surface object and distance transform XTension plugin. By averaging the regression slopes and intercepts across all fit nuclei in a stack, we were able to solve for the parameters of interest. S5B shows the nonlinear regression of the data, based on the model (13) where the choice 0.6 was based on the observation of a breakdown in segmentation at thresholds below 0.6, which approach the level of the most common background noise after background subtraction (S1 Data, tab A). A fit to the six test files provided, spanning the available intensities provides the empirical formula 14 used for all subsequent threshold calculations for tumor nuclei.

(14)

Once this threshold was calculated, the masked tumor DNA were fit using cell body detection with the calculated threshold and 0.3 μm smoothing and 5 μm radius for background subtraction. To account for cavities that could be created using this method that would under-estimate the DNA amount, an additional step to fill holes was included (S3B Fig, panels 7.-8). The preliminary segmentation described above was used to mask a white channel, and these objects were subjected to the fill-holes binary operation slice-by-slice using FIJI. The resulting solid objects were imported back into Imaris, and the binary, filled tumor nuclei were fit to an isosurface with 0.3 μm smoothing and automatic settings. S3B Fig, panel 9. shows that this second isosurface is only slightly larger than that of the empirically thresholded original, except for the filled cavities which represent a much more significant volume change (blue vs. white surface). To provide normal controls, normal nuclei were fit using identical settings, except that they were automatically split with seed points of diameter 3 μm. The normal nuclei were exported from a cell to a surface object for curation, in which normal nuclei were filtered for a distance greater than 4 μm from an edge and lack of CK7 staining. At least 20 normal nuclei were measured in order to estimate the mode of the normal DNA as a control for tumor dna amount, with the modes taken from a Gaussian fit to the DNA histograms. All statistics for tumor cells and normal DNA were exported as.csv files, and these data and derived parameters are included in supplemental spreadsheet 1.

Gentle dissociation of live and fixed human tumor specimens and labeling for flow cytometry and sorting

All tumor specimens were kindly provided by TTAB. For live tumor resections, material not required for pathological analysis was weighed and within 4 hours of removal placed in RPMI media at 4°C. Except for the following changes, gentle dissociation procedures were as described in [65] which shows that the enzyme cocktail used preserves the CD45 receptors present on blood cells. To minimize aggregation and clustering of cells further, we supplemented the cocktail with 0.3 mg/mL hyaluronidase (Worthington). 25 mL enzyme solution per g tissue was used, with an incubation period of 2 h 20 minutes and a shaker speed of 95 rpm. Post incubation, mechanical dissociation was performed with 25 expulsions of the 10 mL serological pipette at the most acute angle possible with the wall of the falcon tube.

All centrifugation steps utilized a Sorvall Legend RT centrifuge employing a swinging-bucket rotor at 1400 rpm for 5 minutes. After filtering the resulting digest with a 70 μm cell strainer, cells were resuspended in 25 mL of 1X RBC lysis buffer at RT for 9 minutes. The reaction was quenched by adding 25 mL PBS, the reaction gently inverted 3–4 times, volumes split into two 25 mL aliquots and cells were resuspended in 1–3 mL PBS. Cells were stained with 1:100 Zombie Red (BioLegend 423109) in the dark in PBS for 10 minutes, then resuspended in PBS for up to 10 minutes. To count cells, a sample was taken on a hemacytometer and cells were imaged on a spinning disk confocal with bright field and 561 nm excitation channels. At this stage the experiment was stopped if >30% dead cells were observed in the red channel.

For pellet-free fixation, cells were combined, in order, with volumes of cold Fluorescence-activated Cell Sorting buffer (FACS, PBS, 0.5% BSA, 2 mM EDTA), PBS, and PBS containing 10% formaldehyde to achieve a final concentration of 2% formaldehyde, 0.1% BSA, 0.4 mM EDTA, and less than 10 million cells/mL. Immediately on the last addition of formaldehyde at 10% concentration, the tube was capped and gently inverted continuously while being placed on a nutator. The fixation reaction was allowed to complete with nutation in the dark at RT for 15 minutes. To quench formaldehyde, we then added an equal volume of 150 mM glycine at pH 7.2 [66]. The quenching reaction was capped and gently inverted, then rocked on a nutator for 10 minutes in the dark at RT. In anticipation of RNA sequencing attempts, we added 1 tube of Superase In RNase inhibitor (ThermoFisher Scientific) to 2.5 mL of FACS buffer at 4°C, and washed cells twice with this solution before washing cells twice with cold FACS. In the final volume appropriate for MACS cells were resuspended in FACS containing this RNase inhibitor.

Cells were subjected to MACS purification using human CD45 beads and LS columns for depletion (Miltenyi Biotec) according to the manufacturer’s instructions, except for the following changes. During CD45 bead incubation, after 10 minutes a mouse anti-EPCAM-AlexaFluor488 antibody (BioLegend, Table 6) was added at 1:50 along with 1:500 DRAQ5 (BD Biosciences 564902). The tube was flicked 6–10 times until bluish throughout and incubated on ice for an additional 10 minutes. 1–2 mL of FACS buffer was added and the labeled cells centrifuged at 300 g for 10 minutes on a Sorvall Legend RT prep centrifuge with a swining-arm rotor.

The supernatant was completely removed and FACS was added to achieve a concentration no higher than 50 million cells per mL. A 40 μL sample of this parent population was saved for later imaging and a portion of this diluted with FACS to measure intensity histograms and calibrate sorting gates (Fig 5A). The supernatant of the parent digest was inspected visually for large particles and if necessary cells were strained again with 70 μm cell strainer.

MACS LS was performed as instructed by the manufacturer except that after a single wash with FACS, a plunge of the CD45 positive cells was performed with 5 mL of FACS and immediately used to load a second equilibrated column. This second column was washed twice with FACS and both CD45 negative fractions were combined and the CD45-positive plunge collected. 40–100 μL samples of CD45-negative and CD45-positive fractions were saved for later imaging. Samples saved for imaging were stored at 4°C for up to six days.

CD45-positive and–negative fractions were concentrated if necessary on a centrifuge.

Sorting epithelial tumor cells by size using the Aria sorter via Live/Dead, EPCAM, DRAQ5, and SSC-W gating

All flow cytometry and sorting was performed using a BD FACS Aria II SORP with violet, blue, green, and red solid-state lasers and 14 color parameters, controlled by BD FACSDiva software. Laser calibration of fluorophore gates was performed using the parent sample, and laser lines employed were the 515/20 (495–525 nm), 610/20 (600–630 nm), and 660/20 (650–670 nm). The FSC and SSC lasers were calibrated in a standard fashion to enable maximal coverage of the data and prevent saturation of the upper limit of the SSC-H vs SSC-W plot (appropriate calibration is shown in Fig 5B). Compensation was not required due to high separation of the fluorophores employed, AlexaFluor-488 (EPCAM), Zombie Red (Live/Dead), and DRAQ5 (DNA). Gates were defined exactly in the manner shown in Fig 5A employing the signal histograms. The Live/Dead gate was drawn using the parent digest, with the upper limit intersecting the left slope of the positive signal at ~FWHM. The EPCAM gate was drawn using the CD45-negative sample, and intersected the valley between the obviously enriched positive and negative signals. The DNA gate was drawn using the CD45-negative fraction, such that the upper limit of the low-DNA gate was set to 1.5X the mode of the DNA signal. This was observed to mark a “shelf” suggestive of polyploid cells, and was settled on because higher levels allowed the appearance of doublets and higher order aggregates in the imaging data.

Size sorting of tumor cells employs the SSC-H vs SSC-W plot with ungated CD45- cells. Size gates were developed using an empirical curve based on the measured area of cells in imaging from gates, and the boundaries used produced optimal sorting (Fig 5 and S16 Fig). SSC-H boundaries were trimmed from 50–200 based on preliminary observations of aggregates arising from those regions, and size gate ranges are as shown in Fig 5D (G1 =“Smallest”, 64–79, G2 =“Small”, 84–99, G3 =“Large”, 101–113, G4 =“Largest”, 114+). After successful DNA gating, gate G4 contained sparing aggregates, cells with protein attached, and very usual cell shapes with very low yield (the remainder taken by subtracting the fractions G1-G3 in S17 Fig from 1). If DNA gating was unsuccessful, cell aggregates became obvious in G3 and rare giant superpolyploid cells could sometimes be observed in G4. We found overall that developing a figure like S16 Fig and quantifying aggregates was essential to calibrating our size sorting, as inspired by [63].

Time-course videos of gentle digestion of FFPE extracellular material to isolate AT2 processes

Distant control and tumor FFPE blocks from a patient diagnosed with a highly invasive LA were provided by TTAB and was processed with IF as described, up to the point right before dehydration. Instead of methanol, tissue sections were immersed in custom round imaging wells 3.75 mm deep and filled with a 180 μL volume containing 1 part PBS and 1 part 1 mg/mL collagenase/dispase preparation (0.5 mg/mL final total enzyme content, Roche, 10269638001), also containing 1 mg/mL hyaluronidase (MolBio). To visualize DNA, 20 μL of 20x Hoescht 33342 was added at time zero. Reactions were allowed to proceed at room temperature for 100 minutes before imaging on a spinning-disk confocal microscope with a 60X/water objective at various times over a period of six additional hours. To prevent evaporation and inhibit any spurious protease activity upon processes, a different version of FACS buffer from that used with MACS (PBS, PH 7.6, 3 mM EDTA, 2% BSA) containing complete EDTA-free protease inhibitor (Roche) was added every two hours in dropwise quantities of less than 50 μL sufficient to maintain fluid level. To provide very gentle loosening of sections, upon addition of protease inhibitor/FACS a pipette tip was very gently moved back and forth in the liquid of the well at a position at least 2–3 mm from the tissue section. S6–S13 Videos were created using Bitplane Imaris^™ in the 3D perspective view after background subtraction as described and after autoscaling of all channels. The small scale bars shown in the videos are for the point closest to the viewer in the perspective.

Spinning-disk confocal imaging of flow cytometry digests and FFPE digests

Flow cytometry digests and gentle digestion kinetics were performed at room temperature with an Olympus IX81 inverted microscope with 405, 488, and 561 laser excitation lines, employing an Andor iXon3 EMCCD camera and controlled by Metamorph software. Flow cytometry digests and resultant sorted cells from gates G1-G4 were imaged with a 20x/air objective. Gentle digest kinetics were imaged with a 60x/water objective with 0.5 or 1 μm z-steps and varying time steps of from one second to one minute to document essential features of the phenomena shown in S6–S13 Videos.

2D analysis of long-axes of Ki-67-staining nuclei

Immunofluorescence, clearing, and imaging was performed as described above, except that imaging was at 25x magnification with a z-step of at most 2 μm and an imaging depth of greater than 200 μm. The rabbit marker used in this case was the highly specific marker Ki-67 marker for active DNA synthesis, as listed in Table 6 [91]. Tumor regions were selected for imaging if they possessed positive CK7 staining in a majority of the imaging volume and if they contained more than one of the developmental morphologies defined in the results. For example, if sheet and sparse (Fig 4) were observed in the same region this region was preferentially imaged over a region with only sheet. Sixteen image stacks were collected spanning eight from the EGFR group and eight from the KRAS group and stages 1–3 represented. The resulting images of nuclei and proliferation markers were segmented in order to mask irrelevant signals using Bitplane Imaris^™ 9.2.1. The following parameters were used for channel 3, stained with TO-PRO-3. Cell body detection, Smoothing = 2 μm, Background Subtraction Sphere Diameter = 5 μm, Cell Threshold = 7.00, Split by Seed Points = 7, Filter by Quality selected manually with the lower threshold at the maximum of the displayed histogram. The resulting Imaris cell object segmentation was exported as a Surface object and filtered for all surfaces containing a sphericity > 0.6. Nuclear masks were also excluded if they were within 10 μm of the edge of the imaging volume.

To identify CK7-positive nuclei, a CK7 surface with default smoothing (.756 μm), Background Subtraction checked, Diameter of Largest Sphere which fits into the Object determined by measurement in the slice view, and lower threshold of 3 units was generated and made 50% transparent. Elements of this CK7 staining visual guide were excluded by a filter if they were of a smaller volume than 600 μm³, the typical volume of a wild-type epithelial type II cell, or they were manually deleted if they appeared to be disconnected from the major CK7 regions and were judged to be irregular elements of autofluorescence artefacts. A filter was applied to the nuclear Surface object for total CK7-staining units and the threshold of this filter was adjusted until nuclei were highlighted only within the CK7-thresholded region. Using this procedure, nuclei below 3 x 10⁴ total units of CK7 signal were always observed to be outside the envelope defined as the CK7-region eliminated from analysis. In some cases single nuclei were identified as non-CK7 by eye and deleted manually. The remaining nuclear surfaces were defined as cancer nuclei, and these were used to mask both the DNA channel and the Ki67 channel. The masked channel 1, which represented both proliferating and non-proliferating cancer cell nuclei was segmented using a cell object with settings of Cell Body Detection, Smoothing Filter Width = 1 μm, Background Subtraction not checked, threshold 20, and split by seed points = 7 μm, seed point quality = default, with only Cell objects with total Ki-67 > 2 x 10⁴, sphericity >0.5, and volume > 150 μm³ retained. These highly actively synthesizing nuclei were defined as the most stringently proliferating cancer nuclei and a surface object based on this cell object was used to mask channel 1 again. This channel masked to include only the highest Ki67 signals was exported to a.tif file along with the DNA channel masked only for the property of being CK7-positive. This 2-channel.tif was opened in FIJI and converted to a hyperstack, and the lookup tables changed such that the DNA channel was displayed as red and the Ki67 channel displayed as green. the stack was made into a color composite and converted to an RGB image. A set of integrated z-projections were created encapsulating each Ki-67 positive nucleus, and the cross-sectional area and long and short axes were measured using the polygon and line tools. Representative z-projections of the nuclei measured along their long axes are shown in S23A Fig.

Estimation of fibrillarin and SON amount and nuclear volumes in human and mouse lung epithelia

Immunofluorescence was performed as described, with channel 1 stains for the appropriate markers applied using conditions shown in Table 7. Segmentation of nuclear volumes for the depictions shown in Figs 7E, 7G and 8A were performed using the Bitplane imaris^™ surface tool using straightforward thresholding approaches to generate isovolume surfaces for nuclei, followed by splitting using the surface scissors tool to separate any closely apposed nuclei. Settings were as described above for cell segmentations, except the isovolume threshold chosen was 1 in all cases.

Analysis of DNA enrichment near the envelope of tumor nuclei

Representative files with tumor cells segmented as described, and from each of the listed patients in the legend of S29 Fig was used to create an eroded set of nuclear volumes using the distance transform tool, with 1.5 μm distance erosion. This surface, representing the core of the nuclei was associated with the original, non-eroded nuclear segmentations. To calculate shell enrichment, the total amount of DNA in the core was subtracted from the total DNA in the nucleus, and the concentration of DNA in the shell was related to the concentration of DNA in the nucleus as a whole. The enrichment shown in S29 Fig is the ratio of those two densities.

Plotting software, general figure notes, and figure graphical enhancements

Figs 1A, 2A, 3A, 4A and 9B, Model 1, were created in Adobe Illustrator. Fig 2B histogram bin regions were colored in red and green for clarity using Adobe Illustrator with the paint bucket tool. Fig 7A was recolored separately in Adobe Illustrator and overlaid for clarity. Scatterplots and histograms were plotted using Igor Pro 6.35A. Violin diagrams in S9 Fig were plotted using Graphpad Prism^™ 8.0. Fig 5A–5C and S17 Fig were created using FlowJo 10, and Fig 5A was colored-in using the Adobe Illustrator Paint Bucket. Boundary lines calculated from 2D segmentations in S2B Fig were calculated using a custom MATLAB script and displayed using Bitplane Imaris^™ in the slicer view. All images shown of normal and tumor lung were captured using Bitplane Imaris^™ using either the 3D view in orthogonal mode, for the case of z-projections, or using the slicer view for 2D depictions of single slices.

Statistics and quantification of percent abnormal cells

Mean, standard deviation, interquartile range and maximum values for volume data were calculated using Igor Pro 6.35A for histograms and plots, or with Excel for supplemental spreadsheet 1. Histogram bins were calculated using the Freedman-Diaconis rule, where bin-width = 2 x IQR/N^⅓. Non-parametric ANOVA and Mann-Whitney difference statistics for S9 Fig were calculated using Graphpad Prism^™ 8.0 with the significance values shown, where all “***** entries have p-values <0.0001. The standard deviation of the fold-change was estimated using the Taylor approximation, assuming independent variables, (15)

Where X and Y correspond, respectively, to the tumor and normal mean parameters and s is the sample standard deviation. The error in the fold change is reported as standard error, approximated as, (16)

Categories of subproportional, proportional or supraproportional cells, were classified simply according to their ratios of their cell model volumes to estimated ploidy. These CV/n values were compared to the μ+σ and μ-σ cutoffs for Normal AT2 cells, assuming 2n ploidy, from the first Gaussian fit of the mixture model (S1 Data, Tumor Cell Data, columns AI and AM, related to Model statistics and fitting tab, Cell V56 ± Cell W56/ see below). Thus, if CV/n < (582–127)/2 = 227, then a cell was called subproportional, and this is the slope of the lower dotted line in Fig 3A, S13 and S14 Figs. Similarly if CV/n > (582+127)/2 then a cell s called supraproportional.

Categories of subproportional, proportional, or supraproportional for patients for Fig 3B were assessed by a conservative compound statistical test defined as follows and illustrated in columns Q-Z of the Patient statistics tab of S1 Data. A chi-square test for the number of cells observed in each category of S13 and S14 Figs against the null distribution of uniformly distributed cells was performed (Columns Q-W of Patient statistics). If the null hypothesis was retained (even distribution) with α = 0.05, then a patient was categorized overall as “Proportional”. If the null hypothesis was rejected (non-uniform distribution), then the supraproportional cells were compared to the subproportional cells using a binomial test with probability 0.5 and α = 0.05 (Column X of Patient statistics). If the null hypothesis was retained a second time (patient 4854 was the only case of this), the patient was categorized as “Proportional”. Otherwise, the patient was categorized as “Abnormal” and potentially in the supraproportional or subproportional categories. If neither the supraproportional or subproportional bin was 2-fold enriched over the other (Columns Z and AA of Patient statistics), then the patient was still categorized as proportional. Otherwise, the patient was categorized simply according to the cell bin with the highest number of cells. For example, if a patient has a non-uniform distribution of cells, and the number of cells in suprapraportional cell bin was 3-fold enriched over subproportional cells, the patient was called “Supraproportional”. On the other hand, if that patient only had 1.35 enrichment in the supraproportional bin, it was still called proportional. See also, S1 Data, Patientwise statistics, columns Q-AB, where cyan cells indicate progression through the decision rule.

Least-squares regression fits to distributions of cell and nuclear volume and N:C ratio used three competing models, all three are invoked only as empirical models for fits to tumor cells, as we assume a host or ensemble of cells living in a community. For normal AT2 cells, we assume a model with two populations with Gaussian distributions related to 2n and 4n.

Regression model 1, two-Gaussian mixture, empirical except for normal cells (17)

Regression model 2, Single Gaussian, empirical (18)

Regression model 3, Lognormal, empirical (19) where the baseline was fixed to zero in the regression. The choice of each model is justified using the rationale briefly recorded in S1 Data, and initial guesses are also recorded therein. We report the parameters from the regression and use the 2n populations as the normal control Gaussians for our main figures (S1 Data). The width parameters from these Gaussians (S1 Data, Model statistics and fitting tab, Cells W56-W57) are used to calculate σ as, (20)

For example, 179 was the fitted value of width1 for the first gaussian peak of AT2 normal cell volumes under the 2-gaussian mixture model (S1 Data, Model statistics and fitting tab, cell W56), which gives σ = 179/1.414 = 127. This value is used to formulate the upper and lower bounds of proportional cells for Fig 3A and 3B and for classification of patients in S13 and S14 Figs.

For quantification of percent abnormal cells, an image of a fitted curve was generated and the areas not within the regions specified by the normal cell gaussian were selected using the magic wand tool in FIJI, and measured against the total area under the curve. To clarify analysis, we approximate the standard error of the curve by that traditionally used for that of a proportion, (21) i.e., as though we could actually count cells that were outside the normal gaussian, where n is the number of cells in each stage (Table 4).

Alignment of mouse and human Pulmonary Surfactant-Associated Protein C N-terminal sequences

Amino acids 1–100 from sequences with Uniprot ID’s P21841, MDMSSKEVLMESPPDYSAGPRSQFRIPCCPVHLKRLLIVVVVVVLVVVVIVGALLMGLHMSQKHTEMVLEMSIGAPETQKRLAPSERADTIATFSIGSTG, and P11686, MDVGSKEVLMESPPDYSAAPRGRFGIPCCPVHLKRLLIVVVVVVLIVVVIVGALLMGLHMSQKHTEMVLEMSIGAPEAQQRLALSEHLVTTATFSIGSTG, were aligned using the Needleman-Wunsch Global Alignment Tool at NCBI (https://blast.ncbi.nlm.nih.gov/) with the default parameters Existence: 11, Extension: 1.

Rationales for selection of images shown in the figures

Normal distant controls: These cells displayed no noticeable trends with genotype or stage of the patients with respect to cell or nuclear size or staining phenotype. A perusal of S1 Data, tab C across patient IDs should convince the reader that normal distant controls are highly invariant in terms of scale. Any images of normal distant control samples were selected based on the ease of viewing particular features or with respect to their ability to provide high contrast and detailed depiction of clearly-staining features such as cell bodies or processes. For normal cells, CK7 staining was assessed for its ability to display alveolar physiology (e.g., Figs 1A and 6).

Fig 1C: The tumor sample shown possesses many features we commonly expected to see when seeing a sample for the first time, including granular TO-PRO 3 staining, fibrous-looking red CK7 staining, nuclei about the size of normal AT2 cells, stromal cells nearby, and some process-like staining with proSPC at cell boundaries.

Fig 1D: The tumor sample shown was selected due to the fact that its flat structure greatly facilitates depiction of the cell boundary simulation by our machine learning algorithm and it allows the reader to gauge for themselves the accuracy of the overall method. Furthermore, it was used during early development of the machine learning algorithm as a testbed and displays a sheet morphology present in about half of our image stacks.

Fig 3D: Cells were selected with a volume:ploidy proportionality and X,Y values close to the center of the datasets shown in 3B. 2D images were selected to display both the cell of interest and their neighborhood and do not show the widest cross-section, showing that the 3D segmentation provides a more accurate depiction of the cell size than the 2D view.

Fig 4A: Growth types shown were observed in tumors regardless of patient genotype, and were selected due to the clarity with which the sheet, solid, and sparse behaviors could be distinguished, as well as the staining quality. Features were also very carefully selected to give the reader a greater sense of our authentic experience of discovery at the microscope. These include the elongated nuclei of the stroma in the sheet depiction, the metaphase nuclei seen in the solid depiction strongly suggestive of proliferation and vitality of the tumor cells, and the doublet nuclei with inclusions seen in the sparse depiction.

Fig 5E: Spinning disk microscope images were first filtered for their depiction of easily distinguished round cells with good brightfield contrast. Cells within these “pretty” images were measured and the mean cell area within a given image was matched as closely as possible to the mode of the peaks shown in the bottom three rows of Fig 5D, Tumor.

Fig 6: RAGE, Tumor was chosen due to its depiction of a tumor-like phenotype, easily distinguished from the Normal alveoli structure above it. Neither genotype of tumor displayed any RAGE expression. proSPC, tumor were selected to display the complexity of the process structure and the greatest diversity of the proSPC staining within the tumor structures for both genotypes. We felt this was especially important because this aspect of human lung physiology has never been reported before.

Fig 7B, 7D and 7F: Figures were chosen for their capacity to convey the specific features being discussed or studied. 7B was selected because it shows that stromal cells have the same protein distribution as giant cancer cells. 7D was selected because it adequately displayed nucleolar staining behavior and was typical of such images. 7F was selected because it displayed speckle-like structure typical of such images.

Fig 8A: Features shown were observed in both genotypes and were chosen as idealized examples to demonstrate existence of the phenomena as they are depicted, and for ease of reference for those wishing to follow-up on these observations.