Project-specific notebooks for CZI Hosted Projects (formerly Tier 2) and CZI Maintained Projects (formerly Tier 1) are embedded in the site

[pablo-gar]

Story: I am a computational biologist who wants to get Census integrated embeddings and their models to perform downstream analysis in my platform and toolkit of choice, for example scanpy in Python

Substory: I know different integration approaches have different properties, and I would like to gain a high-level understanding of the analytical potential that each different integration has. I can follow analytical examples to ground my understanding. For example:

• For any given data, can I predict cell types or diseases? Or any other states?
• Can I map my own data to the embedding space?
• Do the embeddings encode any biological and/or technical variance?
• Can I do in silico experimentation with the models to find candidate genes to go from one state into another?
• Can I perform differential gene expression?
• Can I visualize cells in a 2D integrated scatter plot?
• Can I perform graph-based analysis on the embedding space to learn new biology?
• Can the embeddings map cells from different organisms?
• Etc

Product requirement: Project-specific notebooks for CZI Hosted Projects (formerly Tier 2) and CZI Maintained Projects (formerly Tier 1) are embedding in the site

High-priority demonstations

1. Visualizing a subset of cells in a 2D scatter map
2. Analysis on the neighbors graph for interesting findings.
3. Projecting any cells (census and non-census) into the embedding space.
4. Inference workflow with the actual model (not the embeddings):
   5. e.g. Cell type
   6. e.g. in-silico experimentation

[atarashansky]

Some notes for me

Model candidates: Tier 2: Jure's model, scGPT, Tim NMF factors Tier 1: SCVI, GeneFormer

GeneFormer + SCVI are the only ones with embeddings available atm. Data will need to be subsetted in the tutorial notebooks.

• notebooks with embeddings visualized
• Graph-based analysis on embeddings - leiden clustering, trajectory analysis
• Projecting new data onto embeddings (zero-shot models only)
• Inference workflows (predictive annotation)
  • Vanilla cell type prediction
  • Gene perturbation

Prioritize Tier 1, then move on to Tier 2.

[dsadgat]

@atarashansky could you create tickets here please? Ty!

[dsadgat]

@atarashansky could you please close the completed tickets? And perhaps close this epic (as its completed, correct?) TY!