Feat: Kinetic Regime Clustering (scv.tl.kinetic_clusters)

[beyzoskaya]

New

This PR implements scv.tl.kinetic_clusters, a tool that clusters genes based on their recovered phase-portrait parameters ($\alpha$, $\beta$, $\gamma$) rather than standard expression levels. It also includes:

• A projection method (score_kinetic_clusters) to visualize where these regimes are active on the cell manifold.
• Updated plotting (scv.pl.kinetic_clusters) with log-scale parameter plots and UMAP projection modes.

Scientific Motivation

Genes with similar mean expression levels can have vastly different kinetic drivers.

• High $\alpha$ / High $\gamma$: Indicates a rapid-response/transient regime.
• Low $\alpha$ / Low $\gamma$: Indicates stability/housekeeping.

Standard expression clustering often misses transient driver genes because their average expression is low. This tool allows users to isolate these functional groups based on the differential equation parameters learned by recover_dynamics.

Validation

I benchmarked this method against standard expression clustering on the Pancreas dataset.

1. Metric: Kinetic clustering identified genes with distinct physical properties that had < 5% overlap with standard highly expressed genes.
2. Biological Proof (See Plot below):
   • Kinetic Regime 2 (High Transcription, Moderate Degradation) specifically highlights the Ductal/Progenitor bridge.
   • This confirms the method identifies transient differentiation drivers that are distinct from stem maintenance (Regime 0) or terminal cell markers (Regime 1).

Related issues

None. (New scientific feature proposal).

Checklist

• Code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally