[ENH] Cell type identification

[PrimozGodec]

Implement a widget that annotates a 2D projection of cell types. It operates in multiple modular steps:

1. Assign cell type significance scores to each cell. This is computed only once and must be efficient. This module will be implemented in https://github.com/biolab/orange3-bioinformatics/pull/133
2. Compute a 2D projection (default: t-SNE, use optional Preprocessor).
3. Find regions on the 2D projection (widget use DBSCAN) that are enriched with certain cell types. Allow modifying resolution. The function that will find regions and label them is implemented in https://github.com/biolab/orange3-bioinformatics/pull/134. Every group will be circled with the concave hull. This functionality will be also implemented in https://github.com/biolab/orange3-bioinformatics/pull/134.
4. Visualize regions with the clearest enrichment of cell types.

Workflow

Inputs:

• Data (gene expression matrix)
• Markers (table linking genes to cell types, the output of Marker genes)
• Preprocessor (optional; t-SNE, MDS, PCA, ...)

Outputs:

• Data (gene expression matrix augmented with cell type scores and projection coordinates as meta values)
• Selected Data (selected gene expression matrix augmented with cell type scores and projection coordinates as meta values)

Widget

The widget will be similar to the scatter plot. It will include the control area and the visualization area.

Control area

It includes three main sections:

• Infobox at the top
• Annotation box with widget specific settings
  • epsilon - the setting which will be passed to DBSCAN clustering. The decision is that the user can set this value manually. And reset manual selection with a reset button. After the selection is reset the button is deactivated. When a user does not provide its own setting for epsilon, it is set using a specific formula which will be implemented in https://github.com/biolab/orange3-bioinformatics/pull/134.
  • p-value threshold - the threshold value which is passed to the annotation module. The value is between 0 and 1.
  • show hull - checkbox which shows and hides the hull around clusters. Default: True
  • show labels - checkbox which shows and hides labels for the cell type from the graph. Default: True
  • num of labels - number of labels shown per cluster. Default: 1. Allow up to three.
• Other scatter plot specific settings.

Visualization area

This part is a scatter plot like a graph with additional elements:

• Each cluster is enclosed with a hull.
• Each cluster has annotations somewhere in the middle of the cluster. The number of annotations per cluster is specified in the control area.
• Each cluster has a tooltip (on mouse over) which shows all annotations and their scores.

Widget sketch:

Examples of good visualizations:

[BlazZupan]

Images above are nice. I propose a combination of outlining with dotted line and labelling, like in the bottom figure above. Also, a recently presented paper on "LabelTransfer - Integrating Static and Dynamic Label Representation for Focus+Context Text Exploration" by Qi Han from Stuttgart (pdf seems not to be available on the net) proposes the following label design: