Tutorials#
scDLKit now presents its public tutorial surface around four research tasks.
The goal is to make the paper identity obvious without pretending that every task is already implemented at equal maturity.
For a publication-style audit of the published notebook copies, see the tutorial execution status page. That page shows the current static tutorial surface and the last recorded run date embedded in each published notebook.
Main research tasks#
Status: Pilot
Question answered: Can scDLKit already support a credible low-code adaptation story on labeled human data?
Model and task scope:
Current pilot is implemented on the experimental scGPT adaptation path for
labeled human single-cell RNA data. The published main tutorial stays on the
CPU-practical frozen_probe + head quickstart, while the dedicated annotation
benchmark extends the matrix to full fine-tuning plus the PEFT family.
Figure role: Primary supervised adaptation and PEFT comparison figure family.
Status: Planned
Question answered: Can adapted representations preserve biological structure while reducing technical variation across studies or batches?
Model and task scope: Planned as a paper task with a formal integration metric pipeline.
Figure role: Cross-study transfer and batch-mixing figure family.
Status: Planned
Question answered: Can the adapted model recover perturbation-response structure under a unified benchmark interface?
Model and task scope: Planned as a dedicated perturbation benchmark pillar.
Figure role: Response-prediction and low-label efficiency figure family.
Status: Planned
Question answered: Can scDLKit support spatial task adaptation as a real paper pillar rather than an appendix claim?
Model and task scope: Planned around the Nicheformer-facing spatial benchmark story.
Figure role: Spatial qualitative and task-performance figure family.
Supporting workflows#
These are still valuable and still public, but they are not the main paper-task surface.
Audience: Users starting the stable baseline workflow.
Outputs: Latent embedding, report, loss curve, latent UMAP.
Audience: Users who want the Scanpy interpretation layer after the model step.
Outputs: Latent UMAP, Leiden UMAP, markers, downstream report.
Audience:
Users validating whether deeper baselines help beyond PCA.
Outputs: Benchmark CSV, comparison figure, baseline reference UMAPs.
Advanced / appendix workflows#
These remain documented and tested, but they are no longer equal to the four main research task tracks.
Advanced appendix#
Experimental detail appendix#
Maintainer and smoke path#
Reading order#
If you want the current stable baseline path first:
If you want the current research-facing annotation path first: