The problem. Louvain (day 6) is the community-detection method underneath most single-cell clustering. But it has a hidden flaw: it can produce communities that are internally disconnected — a “cluster” that isn’t actually one connected group. In single-cell terms, a cell type could be split or mixed in ways the algorithm silently permits, and nobody notices because the modularity score still looks fine.

The idea. Leiden refines Louvain with a guarantee: it adds a refinement phase that ensures every community it returns is well-connected, and it converges faster to better partitions. Same modularity-optimisation spirit, but with the connectivity defect fixed and improved efficiency on large graphs.

Why it matters. This is a quietly important correction to a tool I use without thinking — Scanpy switched its default clustering from Louvain to Leiden, so my single-cell clusters are probably Leiden clusters. Reading it is a reminder that “standard method” doesn’t mean “flawless,” and that the defaults in my toolkit encode specific, improvable algorithmic choices. It directly upgrades the day-6 Louvain paper.

Verdict. Foundational as the current default for graph clustering in single-cell. A focused, well-motivated improvement rather than a new paradigm — which is exactly why it matters. Read it right after Louvain to see the bug and its fix.