The problem. A single cell is a point in tens of thousands of dimensions (one per gene). You cannot see that, and naive projections (like PCA) preserve global variance but smear the local neighborhoods — the very cell-to-cell similarities you care about. How do you get a 2-D picture that keeps who is near whom?

The idea. t-SNE converts pairwise distances into probabilities: for each point, which others are likely its neighbors. It then arranges points in 2-D so that a similar neighbor-probability structure holds, minimizing the divergence between the high-D and low-D neighbor distributions. Using a heavy-tailed (Student-t) distribution in the low-D space prevents crowding, letting distinct clusters separate cleanly. The result preserves local structure strikingly well.

Why it matters. For years, the single-cell figure was a t-SNE plot colored by cluster — it’s how the field learned to look at its data, and it’s paired with the Louvain clustering from earlier today. It’s also a cautionary tool: distances between clusters and their sizes aren’t meaningful, a trap I’ve learned to respect. Reading the original clarifies exactly what the plot does and doesn’t say.

Verdict. Foundational for visualization, later largely supplanted by UMAP (coming up in the reading list) for speed and better global structure. Its danger is over-interpretation — it’s a lens, not a measurement. Read it to understand what every embedding plot is actually optimizing.