Downstream Analysis Roadmap
Downstream analysis should turn a trajectory result into biological interpretation: which genes drive a transition, which regulators control them, and how perturbations may redirect fate.
Scope
flowchart TD
A["Cafe trajectory"] --> B["Gene trends"]
A --> C["Driver gene ranking"]
A --> D["GRN inference"]
C --> E["Perturbation prediction"]
D --> E
B --> F["Interactive report"]
C --> F
D --> F
E --> F
Priority Modules
| Priority | Module | Purpose |
|---|---|---|
| P0 | branch-specific gene trends | explain expression dynamics along pseudotime |
| P0 | lineage driver genes | identify genes associated with terminal fate choice |
| P1 | GRN wrapper | connect regulators to target genes |
| P1 | cellxgene-cafe display | inspect trajectory and downstream results interactively |
| P2 | perturbation prediction | predict fate shift after gene perturbation |
| P2 | trajectory alignment | compare dynamic programs across datasets |
Gene Trend Analysis
For each gene \(g\) and lineage \(k\):
\[
y_{ig}^{(k)} = s_{g,k}(\tau_i) + \epsilon_i
\]
The first implementation can use simple smoothers or generalized additive models. The important output is a transition-aware table:
| gene | lineage | score | trend_type | peak_time |
|---|---|---|---|---|
Driver Gene Analysis
Driver genes should be ranked per lineage or edge:
\[
D(g, e) =
\left|\operatorname{cor}(x_g, \alpha_e)\right|
\times
\operatorname{specificity}(g, e)
\]
Candidate wrappers:
- CellRank lineage drivers
- Dynamo vector-field regulators
- random forest feature importance from position prediction
- branch-specific differential expression
GRN Inference
Candidate wrappers:
- SCENIC / pySCENIC
- GRNBoost2
- CellOracle
- Dynamo regulatory analysis
Cafe should normalize GRN outputs to:
| regulator | target | lineage_or_edge | weight | sign | evidence |
|---|---|---|---|---|---|
Perturbation Prediction
Perturbation output should be expressed as a fate or transition shift:
\[
\Delta P_{k}^{(g)} =
P(y = k \mid \operatorname{perturb}(g)) - P(y = k)
\]
Candidate tools:
- CellOracle perturbation simulation
- scGen / CPA-style latent perturbation models
- GEARS-like perturbation models
- method-specific vector field perturbation
cellxgene-cafe Connection
The downstream module should export results through the Cafe Result Schema, so cellxgene-cafe can display:
- pseudotime and lineage coloring,
- velocity arrows,
- milestone network overlay,
- driver gene tables,
- gene trends,
- GRN edge tables,
- perturbation effect tables.
Student Deliverables
Milestone 1:
- run one gene-trend analysis on a standard dataset
- run one CellRank-style driver gene analysis
- write a short method survey for GRN and perturbation tools
- export one result into the Cafe result schema
Milestone 2:
- implement
DriverGeneWrapperdesign draft - implement
GRNWrapperdesign draft - create a downstream demo notebook
- add interactive display requirements for cellxgene-cafe
Milestone 3:
- complete one biological case study
- compare driver genes with known markers
- create publication-quality downstream figures