Roadmap#

Current phase#

scDLKit is in a quality-hardening phase rather than a model-zoo expansion phase.

Immediate goals:

  • stabilize the public API and default behaviors

  • lengthen the public tutorials without changing the overall workflow

  • benchmark the toolkit itself on small built-in Scanpy datasets

  • keep the project gene-expression-only until the quality gates stay green

Immediate milestone#

Target: v0.1.2 quality release

Planned deliverables:

  • quickstart and full profiles in the public notebooks

  • explicit internal quality-suite scripts and benchmark summaries

  • PCA as the classical reference baseline in the comparison tutorial

  • regression checks for PBMC latent quality, classification quality, tutorial artifacts, and runtime budgets

Next phase#

Target: v0.1.3 extensibility release

Planned direction:

  • adapter-based support for user-supplied PyTorch modules

  • keep the built-in registry path for bundled baselines

  • show custom-model integration without changing the Scanpy downstream workflow

Later phase#

Target: v0.2.0 application and downstream analysis expansion

Planned direction:

  • deeper downstream tutorials built around the latent embeddings

  • stronger guidance on when to use PCA versus scDLKit baselines

  • reconstruction and denoising sanity-check tutorials

Deferred work#

  • scverse ecosystem submission

  • spatial baseline support

  • multimodal workflows

  • broad framework-style expansion

The main priority remains a trustworthy baseline toolkit rather than a broad single-cell framework.