Computational biologists can identify the most similar Census cells to their own data for downstream analysis, e.g. cell annotation, in silico perturbation, and quality control.

User stories

I am a computational biologist with my own data in H5AD or Seurat and I'd like to know which Census cells are the most similar to each of my own cells. I need to know the level of confidence on the similarity. I'd like to this so that I:
- Can perform cell type annotation.
- Can assess disease state.
- Can do QC on my data based on prior knowledge. For example if I sequenced brain cells I expect to see only brain-related cells.

An API with temporary namefind_similar_cells that takes input data, and returns the n Census cells most similar to each cell in input and their distances.
- Input data format: AnnData (required), Seurat (stretch)
- Output data: [[input_cell_id_0, [soma_joinid_0, ..., soma_joinid_n], [distance_0, ..., distance_n]], [input_cell_id_1 ...]]
- This API must project input data onto a Census integrated embedding. Initial test case is scVI, but the method should be modularized to allow for future interoperability of models.
- Similarity method: TBD by owner of Epic.
An API with temporary name predict_cells that takes/uses find_similar_cells to predict cell metadata on input data using Census' standard metadata categories (e.g. cell_type, disease, tissue, etc.) It provides a confidence score of predictions.
- Input data format: AnnData (required), Seurat (stretch).
- Parameters: cell metadata category to predict
- Output data: [[input_cell_id_0, predicted_category, confidence_score], [input_cell_id_1 ...]]
Besides model projection, all other methods should be runnable at semi-interactive speed 2-10 Mins for ~200K cells (equivalent to medium size dataset)

Product requirements are flexible for preferred methods, should be defined by SME and Eng.
Product requirements are agnostic about using pre-calculated artifacts that benefit product, e.g. indexing via feiss.