Fig 1.
Workflow of the flow cytometry experiment and the software-based analysis.
(1) Raw data were generated by flow cytometry. (2) FCS files were imported into the Cytolution software. Initial quality control included panel check, compensation data check, automatic control of the time channel and removal of cell debris and cell aggregates. In this step, the user is able to manually select live/dead cells based on a viability dye. (3) The third step is data transformation. In this step, the program automatically determines the needed transformation parameters for each dye dependent on the input data. Optionally, the user is able to manually do the transformation, which was not performed for the present dataset. (4) The next steps are cutoff definition and population definition. The program automatically determines multiple cutoff values per dye based on the input data (+/-, low, +/+). Optionally, the user is able to manually shift the cutoffs, which was not performed for the present dataset to obtain an unbiased view on the data. (5) Based on these cutoffs, a population tree can be built by the user. In the present study, we determined the populations CD11b+Ly6G+F4/80- granulocytes, CD11b+Ly6G-F4/80+ macrophages, CD11b+Ly6G+F4/80+ MDSCs, CD19+ B-cells, CD3+ T-cells, CD3+CD8+ cytotoxic T-cells, CD3+CD4+ Thelper-cells as of interest for us. (6) The program calculates multiple subclusters based on AI/machine learning-algorithms. The subclusters can either be independent from the defined population tree or they are classified as subclusters within the pre-defined subpopulations from the population tree. All pre-defined populations and AI-generated subclusters can be visualized in 2D or 3D UMAPs (or t-SNE) based on their expression patterns. Furthermore, absolute cell numbers or ratios can be exported or displayed as graphs in the program itself. The most innovative part is to further explore the AI-based subclusters in the Cluster explorer. In this part of the program, the user is able to see why distinct subclusters were generated based on SSC, FSC and other marker expressions.
Fig 2.
Immune cell populations in the fracture hematoma.
We used the automatically generated cut-off values for our stained surface markers CD11b, Ly6G, F4/80, CD3, CD4, CD8 and CD19 to evaluate the following user pre-defined immune cell populations: A) CD11b+Ly6G+F4/80- granulocytes, B) CD11b+Ly6G-F4/80+ macrophages, C) CD11b+Ly6G+F4/80+ MDSCs, D) CD19+ B-cells, E) CD3+ T-cells, F) CD3+CD8+ cytotoxic T-cells, G) CD3+CD4+ Thelper-cells. H) A 3D UMAP of all selected cell populations for all samples is displayed. I) 3D UMAP of all selected cell populations in the hematoma samples from sham-OVX mice. J) 3D UMAP of all selected cell populations in the hematoma samples from OVX mice.
Fig 3.
Immune cell populations in the bone marrow.
We used the automatically generated cut-off values for our stained surface markers CD11b, Ly6G, F4/80, CD3, CD4, CD8 and CD19 to evaluate the following user pre-defined immune cell populations: A) CD11b+Ly6G+F4/80- granulocytes, B) CD11b+Ly6G-F4/80+ macrophages, C) CD11b+Ly6G+F4/80+ MDSCs, D) CD19+ B-cells, E) CD3+ T-cells, F) CD3+CD8+ cytotoxic T-cells, G) CD3+CD4+ Thelper-cells. H) A 3D UMAP of all selected cell populations for all samples is displayed. I) 3D UMAP of all selected cell populations in the bone marrow samples from sham-OVX mice. J) 3D UMAP of all selected cell populations in the bone marrow samples from OVX mice.
Fig 4.
3D UMAP of immune cell AI-based subclusters based on all samples of the dataset.
To construct the initial high-dimensional graph, UMAP creates what is known as a “fuzzy simplicial complex.” This is essentially a weighted graph, where the weights on the edges represent the probability of two points being connected. To establish these connections, UMAP expands a radius around each point and links points when their radii overlap. The choice of radius is crucial, if it is too small, it results in small, isolated clusters, while a large radius connects everything indiscriminately. UMAP addresses this by selecting a radius locally, based on the distance to each points nearest neighbor. It then introduces “fuzziness” by reducing the probability of connection as the radius increases. Lastly, UMAP ensures that each point is connected to at least its closest neighbor, preserving local structure while balancing it with the global layout. Therefore, a UMAP show similarities between the different clusters after a dimensionality reduction-.
Fig 5.
Only subclusters which were significantly different between sham-OVX and OVX mice in the hematoma are displayed. Data for all other subclusters are displayed in S1 and S2 Tables. A) “B-cell subcluster31” in the hematoma and B) the bone marrow. C) “Granulocyte_subcluster0” in the hematoma and D) the bone marrow. E) A 3D UMAP of all B-cell subclusters is displayed. F) A 3D UMAP of all granulocyte subclusters is displayed.
Fig 6.
Cluster explorer analysis for granulocyte subclusters.
We further looked at the unique features of the 27 granulocyte subclusters. A) FCS vs. SSC for all subclusters and B) only for subcluster_0. C) 7AAD (dead staining) vs. Ly6G for all subclusters and D) only for subcluster_0. E) CD11b expression for all subclusters and F) only for subcluster_0. Automatically determined cut-off values can be seen in the graphs.