Mapping mammary gland architecture using multi-scale in situ analysis

Abstract

We have built a novel computational microscopy platform that integrates image acquisition, storage, processing and analysis to study cell populations in situ. This platform allows high-content studies where multiple features are measured and linked at multiple scales. We used this approach to study the cellular composition and architecture of the mouse mammary gland by quantitatively tracking the distribution and type, position, proliferative state, and hormone receptor status of epithelial cells that incorporated bromodeoxyuridine while undergoing DNA synthesis during puberty and retained this label in the adult gland as a function of tissue structure. Immunofluorescence was used to identify label-retaining cells, as well as epithelial cells expressing the proteins progesterone receptor and P63. Only 3.6% of luminal cells were label-retaining cells, the majority of which did not express the progesterone receptor. Multi-scale in situ analysis revealed that luminal label-retaining cells have a distinct nuclear morphology, are enriched 3.4-fold in large ducts, and are distributed asymmetrically across the tissue. We postulated that LRC enriched in the ventral mammary gland represent progenitor cells. Epithelial cells isolated from the ventral versus the dorsal portion of the gland were enriched for the putative stem cell markers CD24 and CD49f as measured by fluorescence activated cell sorting. Thus, quantitative analysis of the cellular composition of the mammary epithelium across spatial scales identified a previously unrecognized architecture in which the ventral-most, large ducts contain a reservoir of undifferentiated, putative stem cells.

Fig. 1

LRC are a heterogeneous population. (A) Section through a mammary duct where LRC and myoepithelial cells have been labeled using fluorescent markers. The image was background-corrected to improve visualization. Open arrowheads indicate the position of luminal LRC, an arrow points at a suprabasal LRC and white arrowheads show the location of myoepithelial LRC. Saturated green areas in the stroma correspond to non-specific staining. They can be identified by their yellow, flat appearance and non-nuclear localization versus the punctate nuclear pattern of staining in true P63⁺ cells. Scale bar, 10 µm. (B) LRC frequencies in the luminal and the myoepithelial compartment show that LRC are more abundant among myoepithelial cells (*P = 0.027), possibly reflecting early differentiation of these cells. The number of cells counted was 25 317 and 2975 for the luminal and myoepithelial compartments, respectively, in n=4 animals. (C) At mid-pregnancy, proliferating cells are 59-fold more abundant in the luminal layer than in the myoepithelial layer (**P = 0.0065). The number of cells counted was 10 567 and 6902 for luminal and myoepithelial compartments, respectively, in n = 4 animals.

Fig. 2

Multi-feature image analysis shows that LRC have a distinct nuclear signature. (A) Nuclear size analysis shows that LRC nuclei are smaller than non-LRC (****P =3.34 ± 10⁻⁹) and PR⁺ LRC (***P =0.003) nuclei. The number of nuclei analyzed was n= 226, 321 and 29 for LRC, non-LRC and PR⁺ LRC, respectively. (B, C) LRC have smaller SF and larger NBE than non-LRC (P=7.11 × 10⁻⁶ and P=3.78 × 10^−11, respectively) and PR⁺ LRC nuclei (P = 0.004 and P = 0.003, respectively), as shown by their cumulative distribution functions (CDF). (D) Chromatin texture cumulative distribution functions (CDF) show a different chromatin organization in LRC with respect to non-LRC (P = 3.09 × 10⁻⁹) and PR⁺ LRC (P = 0.025). (E) Nuclear size vs. chromatin texture. The numbers indicate the percentage of cells in each population above and below the line, which can be quantitatively described by eqn (4). The line delineates the bulk of the differentiated and undifferentiated populations. (F) Degree of morphological differentiation for the populations characterized in situ. The number of nuclei analyzed was n = 9, 226, 321 and 29 for P63⁺ suprabasal LRC, all LRC, non-LRC and PR⁺ LRC, respectively.

Fig. 3

Monte Carlo simulations demonstrate that LRC form small clusters in larger ducts. (A) Medium-sized duct where LRC (green nuclei) form small clusters. The duct was also stained for PR (red nuclei) and counterstained with DAPI (blue). The graph connecting neighbor cells used to evaluate the M-function (see Materials and Methods) is overlaid. Scale bar, 10 µm. (B–D) M-Function analyses for LRC in LD, MD and SD (respectively) reveal significant clustering in LD and MD.Mindicates the degree of LRC clustering observed in tissue samples, while U and L show the upper and lower limits for the range of clustering values obtained from simulations of random LRC distributions. M values above U indicate that the degree of clustering of LRC observed in situ cannot be recapitulated in random simulations. Therefore LRC are significantly clustered in LD and MD. In SD, M values remain within the range obtained in simulations (U and L), indicating that the pattern of LRC distribution is likely to be random in smaller ducts. The number of cells analyzed was n = 127, 295 and 18 for LD, MD and SD, respectively.

Fig. 4

LRC do not preferentially associate with the proliferation marker Ki67, but they significantly exclude the differentiation marker PR. (A–C, E–G) M Indicates the degree of association between LRC and Ki67⁺ (A–C) or PR⁺ (E–G) cells. U and L show the upper and lower limits for the range of association values obtained from simulations of independent distributions of LRC and Ki67⁺ (A–C) or PR⁺ (E–G) cells. (A–C) Distribution of Ki67⁺ cells with respect to LRC in LD, MD and SD (respectively) shows that both populations are independently arranged. The number of cells analyzed was n = 220, 334 and 6 for LD, MD and SD, respectively. (D) Dual staining for LRC and Ki67. Nuclei were counterstained with DAPI (blue). Arrow points at a cell where both markers co-localize. Scale bar, 20 µm. (E–G) Distribution of PR⁺ cells with respect to LRC in LD, MD and SD (respectively) shows that retention of BrdU correlates with the absence of PR expression in LD and MD (where M has a minimum below L at r=0, and L is the lower limit of clustering density values obtained in the simulations). At all other distances (r > 1) there is no relationship (aggregation or separation) between LRC and PR⁺ cells. The number of cells analyzed was n=129, 310 and 20 for LD, MD and SD, respectively. (H) Dual staining for LRC and PR. Nuclei were counterstained with DAPI (blue). Arrow points at a cell where both markers co-localize. Scale bar, 20 µm.

Fig. 5

Tissue-wide analysis shows that LRC are more abundant in larger ducts. Relative LRC frequency in each type of duct normalized to the LRC frequency in the entire gland shows LRC enrichment in LD (* P = 0.015) and MD (** P = 0.006). The number of nuclei counted was 7623, 12 551 and 1305 for LD, MD and SD, respectively in n = 4 animals.

Fig. 6

Fluorescence-activated cell sorting reveals an asymmetric distribution of mammary repopulating cells. (A) Ventral tissue was mainly composed of large (blue) and medium (green) ducts, while dorsal tissue was principally formed by medium and small (red) ducts. Ducts were selected and imaged with a homogeneous spatial distribution across n = 4 mammary glands of 18-week-old females. (B) Frequency of CD24⁺CD49f^high cells is higher in ventral tissue than in dorsal tissue (***P=0.0048). (C, D) Density plots representing the cell number in each gated population (ellipses) in dorsal and ventral tissue, respectively, show that CD24⁺CD49f^high cells (red ellipses) are the only population that shows a significant change. Notice how the saturated area within the CD24⁺CD49f^high gate for ventral cells is absent in the plot for dorsal cells. Axes indicate arbitrary intensity units (a. u.). (E) Cartoon representing our model for stem cell distribution along the mouse mammary gland with a reservoir of proliferative potential in larger ducts.