A Worked Reconstruction of a Curved Cell-State Path

Reconstructing One Path in Three Ways

A curved path of cell states can look simple in a two-dimensional embedding and still yield different inferred orders from different methods. A principal curve summarizes a smooth geometric trend, a minimum spanning tree (MST) connects cells with short edges, and diffusion pseudotime (DPT) compares cells through transitions on a neighborhood graph. Each result depends on a different representation of the cells. This worked reconstruction uses the same five hypothetical cells to make those distinctions explicit.

The example is deliberately small enough to calculate by hand. It is not evidence that any particular method is generally preferable. In a real analysis, the fitted curve, the coordinates used to build an MST, and the affinity graph used for DPT may differ, and each choice should be checked for its scientific and computational consequences.

These definitions describe different objects: a curve in a coordinate space, a tree of cell-to-cell edges, and a distance derived from transitions. The earlier material on principal curves, MSTs, and diffusion maps provides the background for these constructions; here the emphasis is on calculating and comparing their outputs. A comparison is meaningful only after stating which coordinates or graph are used, how the root is chosen, and which diffusion time is evaluated.

A Synthetic Curved Path and Its Principal-Curve Order

Consider five hypothetical cells, A through E, represented by two-dimensional coordinates. Suppose a smoothing procedure returns the following idealized fitted principal curve, a unit-circle arc running from angle \(0\) to \(3\pi/2\):

The plotted coordinates are A \(=(1,0)\), B \(=(\sqrt{2}/2,\sqrt{2}/2)\), C \(=(0,1)\), D \(=(-\sqrt{2}/2,\sqrt{2}/2)\), and E \(=(0,-1)\). For this worked example, assume each cell projects to the corresponding point on the stated arc. This is an idealized fitted result: with real data, smoothing, projection, and boundary choices would determine the estimated curve and projections.

A principal curve's self-consistency condition helps explain what the fitted curve is intended to represent: local averages of observations associated with a curve position should lie on the curve. That condition does not by itself guarantee a unique curve, a biologically correct root, or a reliable ordering. The example uses an explicitly stated fitted arc so that the projection and arc-length calculations are reproducible, rather than implying that five points uniquely determine a principal curve.

Building the Cell MST

Now build an MST using straight-line Euclidean distances between the same displayed coordinates. A Euclidean edge length is the square root of the sum of squared coordinate differences. The three consecutive edges A–B, B–C, and C–D each have length \(\sqrt{2-\sqrt{2}}\). The distance A–E is \(\sqrt{2}\), while D–E is longer:

The MST has minimized total edge length, not recovered the assumed arc's direction or length. In particular, an MST does not intrinsically define biological time: its order depends on the root, its branch structure, and the metric used to calculate edges. If coordinates are an embedding that distorts distances, an MST built there can inherit that distortion.

Calculating Diffusion Pseudotime on a Cell Graph

For contrast, suppose a separately specified local-affinity graph connects the cells in a chain A–B–C–D–E, with equal edge weights. This graph might be built from features other than the two displayed coordinates. Define a lazy random walk: at each step the chain stays at its current cell with probability \(1/2\) and uses the remaining probability to move to a neighboring cell, divided equally between neighbors. At an endpoint, the only neighbor receives the remaining probability. This makes the transition matrix \(P\) reversible, with stationary probabilities proportional to node degree: \(\pi=(1/8,1/4,1/4,1/4,1/8)\).

At diffusion time \(1\), define the diffusion distance from root A to cell \(i\) as the stationary-weighted distance between their one-step transition distributions:

The stationary weights matter: differences in transition probability at a cell with small stationary probability receive more weight. This is a standard reversible-graph diffusion-distance convention. Other diffusion times or normalizations can produce different numeric distances, so the convention must accompany reported values.

Why the Orderings Differ

The three reconstructions can be compared directly. The principal curve follows the stated smooth arc and orders cells by projected arc length. The MST uses displayed Euclidean geometry and shortcuts from A to E. DPT uses transition behavior on the separately specified chain graph and places E last. Each method answers a different operational question, as summarized below.

The example's diffusion graph and displayed coordinates intentionally encode different relationships: the graph preserves a local chain, while straight-line distances make E close to A. That contrast isolates the effect of representation. In practice, if all methods are built from the same embedding or affinity matrix, their outputs are not independent checks; shared preprocessing choices can create shared errors.

Diagnostics, Failure Modes, and Limitations

A useful reconstruction includes checks on both the input representation and the resulting order. For a principal curve, inspect cell-to-curve projection distances, the fitted curve's smoothness, and whether projections are stable when the smoothing choices change. Large residuals can indicate that a single curve is a poor summary, while close or looping segments can make projection ambiguous. A curve can also be sensitive to endpoints, which are especially consequential when arc length is measured from a root.

For an MST, inspect unusually long edges, short-circuit edges across nearby branches, and whether small perturbations to coordinates change the tree. An MST is always a spanning tree when the input distances connect all cells, but that mathematical property does not establish that its edges represent plausible developmental transitions. For DPT, check graph connectivity, the root, neighborhood and kernel choices, and whether the ordering is stable across diffusion times. A disconnected graph has no finite diffusion path between components, and weak or spurious edges can dominate apparent connectivity.

The main alternatives are not interchangeable estimators of a single guaranteed true ordering. A curve gives a continuous geometric path but may oversmooth branching structure. An MST makes branches explicit and offers paths through a tree, but it can be sensitive to distance geometry and can impose a tree even when the state structure is not tree-like. Diffusion methods use multistep connectivity and can be less dependent on individual edges, but their output depends on graph construction, transition normalization, root, and diffusion time. None of these methods alone establishes elapsed time, causal direction, or a unique biological lineage.

Common Mistake

A common mistake is to say, “The MST is the shortest path through the cells, so it must give the true cell order.” The MST minimizes the total length of a spanning tree; it does not minimize every root-to-cell path subject to following a presumed curved biological trajectory. In this example it makes a valid minimum spanning tree by connecting E directly to A. The correction is to state the graph, metric, and root, calculate the resulting paths, and interpret them as an ordering under those choices—not as observed time or proven lineage.