title: "Celltype comparison"

author: "Bernd Jagla"

date: "6/7/2018"

runtime: shiny

output:

html_notebook:

fig_caption: yes

number_sections: yes

toc: yes

```{r whichFiles, echo=TRUE}

```{r loadclaire, eval=FALSE, echo=TRUE}

```{r filesDF, eval=FALSE, echo=TRUE}

```{r, eval=FALSE}

```{r, eval=FALSE}

Cell assignements performed by Claire are compared to automatically identified cells by FlowSOM.

Input cell type is the starting cell population for the SOM. The cell type referes to the cells under investigation.

Cell types in FlowSOM are identified by having more than 50% of the cells belonging to that cluster also belonging to a given cell type defined by Claire.

E.g. input: CD8Naivelike, cell type: CD8CD95pos => cc=0.94

vs.

input: CD3pos, cell type: CD8CD95pos => cc = 0.15

```{r summaryTable, eval=FALSE, echo=TRUE}

```{r setupPlotDefaults, echo=TRUE, eval=TRUE}

```{r plot, echo=TRUE, fig.height=10, fig.width=10}

Here, we select the clusters that have more than 50% of cells that Clair found and take all the cells. The correlation is getting better because we are including more cells, but these cells have not been identified by Claire, and there are still other cells that have been identified by Claire but were in clusters with less than 50% of these cells.

```{r plotClusterFSOMSetup, echo=TRUE}

```{r plotClusterFSOM, echo=TRUE, fig.height=10, fig.width=10}

We now look at what the SOM has learned. The MST and FlowSOM two ways of presenting this information for a given SOM. Here, want to compare multiple SOMs together and investigate how the medians of the SOMs correspond to the median of maunally defined clusters.

```{r medianValues, eval=FALSE, echo=TRUE}

The sample ratio defines how many medians should be sampled from the overall number to reduce the display speed. Initially to set up the parameters this should be a low ratio (0.1) and can be set to 1 (everything) once the parameters are set.

We can change the plot order to see if e.g. background points are hidden underneath the flowSOM points.

Individual samples can be removed / added.

```{r prepareData, echo=TRUE}

```{r plot2dSetupDefaultValues, eval=TRUE, echo=TRUE}

select the cell types that visualize their location on the umap.

** Be aware that the sample ratio is in effect for the line plot/Clustering and others.

```{r plot2d, echo=TRUE, fig.height=10, fig.width=10}

```{r linePlot2, fig.width=11.7, fig.height=8.27, message=FALSE, warning=FALSE, echo=TRUE, eval=TRUE}

The reason for this is that the PCA only looks at the variance without taking into account the distances between different points.

```{r, eval=FALSE, echo=TRUE}

The MST is really expensive to calculate and therefore not a good approach here.

```{r, eval=FALSE, echo=TRUE, fig.height=10, fig.width=10}

```{r umapParameters, echo=TRUE, eval=FALSE}

```{r umapPlot, echo=TRUE}

```{r umapCelltypesDefaults, echo=TRUE}

```{r umapCelltypes, echo=TRUE}

I am still not happy with neither the line plots (they are not showing a clear enough separation between cell types) and the Umap clustering. Though in the Umap clustering it is possible to see some different clusters the number of clusters is not sufficient to effectively show different cell cell types based on median values.

What comes to mind now is to apply a dicision tree and some clustering for supervised / unsupervised analysis of the median values.

This might be used to come up with an alternative gating strategy...

Here, we compare the results from a decision tree. The data is split in training (~60%) and test (rest) and the results from the test are shown. We look at the predicions for all samples as well as the samples from donors and recipients analyzed serately.

The clustering is performed on the median vectors. I.e. the trained cluster from the SOM (without Claire's median values.) We don't know the number of cells at this point.

We show the split between donor and recipient because this is supposed to be the biggest source of variation.

Be aware that the ratio defined above also impacts the number of clusters available here.

The goal here is to validate that we can potentially learn how to classify the cells automatically.

```{r c50defaults, echo=TRUE}

```{r c50, echo=TRUE}