The problem. Single-cell analysis splits into two worldviews: clustering says cells fall into discrete types; trajectory inference says they lie on continuous paths. Real tissue is both — distinct types and transitions between them. Forcing one view loses the other, and full trajectory methods can overreach on data that’s mostly discrete.

The idea. PAGA (partition-based graph abstraction) computes a coarse-grained graph where nodes are clusters and edges are weighted by how strongly two clusters are connected in the single-cell neighbour graph. The result is a topology-preserving map: it shows which cell states are linked (and might transition) without committing to a false continuous ordering where none exists. You can then run pseudotime within the connected parts.

Why it matters. PAGA is a standard Scanpy step and a pragmatic resolution of a real tension — it lets me see global structure (which types relate) before deciding whether a continuous trajectory is even warranted. From the same group as Scanpy and scVelo, it reflects a consistent, sober design philosophy: give an honest coarse map rather than an over-confident fine one.

Verdict. Foundational for connecting clustering and trajectories, and a sensible default for surveying structure. Read it as the reconciliation of the field’s two default lenses — discrete and continuous — into one graph.