Common Misreadings of Inferred Pseudotime

What an Inferred Pseudotime Can—and Cannot—Say

Pseudotime is useful because cells can occupy different positions along a biological process even when the experiment did not measure them repeatedly over clock time. An inferred score can arrange cells from a chosen root or starting state toward other states, helping describe patterns in a cross-sectional expression dataset. The interpretation becomes unreliable, however, when the score is treated as a timestamp, when distances along an inferred graph are read as durations, or when a reproducible ordering is taken as proof that only one trajectory is possible.

The distinction is not merely caution about wording. These claims require different information. A ranking requires evidence about relative positions under a chosen model. Elapsed time requires temporal calibration. Biological duration requires information about the rate of progression. A unique trajectory requires evidence that plausible alternative structures are not equally compatible with the observed cells and analysis.

In the earlier material on assigning pseudotime along an MST and on diffusion pseudotime, the score is constructed from a graph or diffusion geometry. The earlier material on resampling stability addresses how much an inferred ordering changes under repeated analyses. Those constructions can support statements about relative arrangement under their assumptions; they do not automatically establish the stronger biological claims discussed here.

Misreading 1: Treating Pseudotime as Clock Time

A pseudotime value such as 2.4 is ordinarily a coordinate, not a statement that a cell was observed 2.4 hours after another cell. Even if a method reports nonnegative values increasing from the root, the numerical scale may be rescaled or transformed without changing the ordering. A cell at score 4 is therefore not necessarily twice as far through a process, or twice as old in that process, as a cell at score 2.

The basic reason is that an ordering contains less information than a clock. If one strictly increasing transformation is applied to every score, the order is preserved, although the gaps between scores can change. The transformed values still describe the same ranking, but their differences cannot be assigned an elapsed-time interpretation merely because they are numerical.

A real experiment may include known collection times. Those times can be used to assess or calibrate an inferred ordering, but their existence does not make every pseudotime score a timestamp. Calibration requires an explicit model relating the score to measured time, appropriate temporal coverage, and an evaluation of how well that relationship fits. If the score is fitted only to expression similarity, it should be reported as a relative coordinate unless such a relationship has been established.

Misreading 2: Treating Graph Distance as Biological Duration

A graph edge length or a path length summarizes geometry under a chosen representation and distance rule. For example, it may reflect separation in a reduced expression space, or accumulated distances along graph edges. It does not measure how long cells take to move between states. Biological duration depends on progression rates, which can differ along a path and between branches.

A simple relationship makes the missing information clear. If \(s\) denotes distance along a path and \(v(s)\) is the rate of progression with respect to clock time, then the time required to traverse a path segment is represented by integrating the reciprocal rate over that segment. Distance alone suffices only under additional rate assumptions.

This formula assumes that a meaningful path coordinate \(s\) and positive progression rate \(v(s)\) have been defined, and that the rate is expressed in path-distance units per unit of clock time. An inferred graph generally supplies neither the calibrated clock units nor the rate function. Even a perfectly measured path length therefore does not, by itself, give \(\Delta t\).

Graph geometry can also change with choices such as the representation, neighborhood scale, or distance metric. Those choices may be reasonable for estimating cell-state relationships, but they should not be mistaken for biological clocks. The earlier material on uneven cell density and trajectory geometry is especially relevant: cell density and geometric arrangement can affect what a graph looks like, without providing a direct measure of how quickly cells traverse it.

Misreading 3: Treating Stable Ordering as Proof of One Unique Trajectory

An ordering that recurs across resamples is evidence of a particular kind of robustness: under the resampling scheme and analysis choices used, the relative ranks of cells or groups of cells change little. It is not, on its own, proof that the biological process follows a single path. A scalar ordering can be stable even when important aspects of the graph’s branching structure, root, or interpretation remain uncertain.

This distinction matters because different trajectory structures can be consistent with similar expression patterns among the cells that were actually sampled. A dataset may have sparse coverage between states, may miss a branch, or may contain states whose relative positions are not resolved by the available measurements. Repeating an analysis on similar resamples cannot guarantee that such unobserved alternatives have been ruled out.

A Short Justification: What an Order Preserves

The monotone-transformation property used above follows directly from the definition of a strictly increasing function. If \(\tau_i < \tau_j\), then for strictly increasing \(g\), \(g(\tau_i) < g(\tau_j)\). Consequently, all pairwise rank comparisons are preserved. But differences need not be preserved: in general, \(g(\tau_j)-g(\tau_i)\) is not equal to \(\tau_j-\tau_i\), nor does it have to be a constant multiple of that difference.

For example, with \(g(\tau)=\tau^2\) on nonnegative values, the gap from 0 to 1 is 1, while the gap from 1 to 2 is \(4-1=3\). Thus, rank stability under a monotone change of scale cannot justify interpreting score gaps as equal durations. This is a mathematical property of the score, not a defect unique to any one trajectory method.

Diagnostics, Comparisons, and Limitations

A useful diagnostic asks whether a proposed interpretation is supported by the quantity actually estimated. To assess ranking, examine rank agreement across reasonable perturbations and report uncertainty where it is substantial. To assess topology, compare plausible graph or branch structures rather than looking only at one scalar score. To assess clock-time claims, compare the inferred ordering with independently measured time points and evaluate an explicit calibration model. To assess duration, seek evidence about rates; geometric distance is not a substitute.

These checks should be interpreted conditionally. Resampling stability describes sensitivity to the resampling and pipeline examined, not every source of uncertainty. A high rank agreement does not establish correct root selection, adequate sampling of transitional states, absence of missing branches, or biological uniqueness. Conversely, an unstable ordering can indicate insufficient information, but it may also reflect a genuinely ambiguous region in which imposing a strict order is not warranted.

The principal alternatives answer different questions. MST path pseudotime and diffusion pseudotime can each provide a root-relative ordering, but their scores arise from different graph or diffusion constructions. A principal curve provides a continuous geometric representation under its own assumptions. Comparing these approaches can reveal method sensitivity; agreement is useful supporting evidence, but none converts an uncalibrated coordinate into time or establishes uniqueness by agreement alone.

A common mistake is to say, “The pseudotime difference is 3, so the cells are three hours apart; the longer graph edge also means the process takes longer there.” This reasoning assigns time units without calibration and assumes a rate from geometry alone. The correction is to report the score as a relative coordinate, identify the graph metric if discussing distance, and reserve duration claims for analyses supported by time measurements or defensible rate estimates.

The limitations are substantive: trajectories inferred from snapshot expression data depend on which cells were sampled, how expression was represented, which root and method were chosen, and whether relevant states are observed. A careful report specifies these choices, separates rank claims from topology claims, and states what independent evidence would be required for temporal or kinetic conclusions. It avoids presenting one inferred path as uniquely established when alternatives remain plausible.