Pseudotime as a Relative Ordering

What Pseudotime Represents

Single-cell RNA sequencing usually measures many cells at one or a few collection times. A trajectory analysis can arrange cells by how their measured expression profiles relate to an inferred progression of cell states. The resulting pseudotime is a coordinate assigned to each cell along that inferred path. It is useful for describing relative position, but the coordinate is not a direct measurement of how long a cell has been progressing.

Let \(i\) index cells and let \(\tau_i\) denote the pseudotime assigned to cell \(i\). If cells \(i\) and \(j\) lie on the same oriented path and \(\tau_i < \tau_j\), the analysis places cell \(i\) earlier than cell \(j\) along that path. This is an ordering claim about inferred cell states. By itself, it does not say that cell \(i\) was observed earlier in calendar time, that it will become cell \(j\), or that the transition takes a particular duration.

The interpretation requires assumptions. The sampled cells must contain enough information to infer a meaningful progression; the path must be a reasonable description of the biological states under study; and the expression differences used by the analysis must reflect relevant state variation rather than being dominated by technical effects or unrelated cell differences. Even if these conditions are plausible, a cross-sectional collection does not directly observe individual cells moving along the path. Pseudotime summarizes an inferred arrangement of cells, not a tracked history of each cell.

There are three distinct choices to keep in mind. The ordering says which cells are placed before or after others. The orientation says which end is called early and which late; it may require a known starting state or another biological anchor. The scale specifies the numerical coordinates assigned along the oriented path. An analysis may establish a useful ordering without establishing either a biologically correct orientation or a time-calibrated scale.

Why the Numerical Scale Is Usually Arbitrary

Suppose an analysis assigns values \(0.2\), \(0.6\), and \(0.9\) to three cells. These numbers encode an order. Unless the method has been calibrated against suitable time measurements under an explicit model, they are not days, hours, or any other duration. A different procedure could assign \(2\), \(6\), and \(9\), or apply a nonlinear increasing transformation, and preserve the same ordering while changing all numerical gaps.

The core reason can be stated directly. Let \(g\) be a strictly increasing function. For any two cells \(i\) and \(j\),

Thus \(g(\tau_i)\) is a different numerical representation of the same strict order. The size of a difference generally changes: \(g(\tau_j)-g(\tau_i)\) need not equal \(\tau_j-\tau_i\), nor be proportional to it. A decreasing transformation reverses the order, so the direction must be fixed before interpreting “earlier” and “later.”

This invariance concerns a transformation of an already assigned set of values. It does not imply that two trajectory methods necessarily infer the same order. If two analyses rank cells differently, applying an increasing transformation to either output cannot eliminate that disagreement. A known root can help orient an order, but it does not, by itself, turn the scale into elapsed time.

Ordering Is Not Elapsed Time

Elapsed time is measured from a specified clock or event, such as hours since stimulation. Pseudotime is inferred from cell-state data. These quantities may be associated: cells collected later after a controlled perturbation might tend to receive later pseudotime values. Association does not make the quantities interchangeable. The association could be imperfect, and the pseudotime scale still needs a justified calibration before differences can be read as durations.

In particular, a pseudotime difference of \(0.2\) does not generally mean half as much progression as a difference of \(0.4\), and it does not mean a fixed number of hours. To interpret a coordinate as time would require additional information and assumptions—for example, measurements tied to known time points and a defensible model linking the inferred coordinate to elapsed time. Even then, the calibration and its uncertainty should be reported rather than silently treating the original coordinate as a clock.

A Justification for Comparing Orders, Not Raw Gaps

A direct comparison of orderings can use ranks. If \(n\) cells are assigned distinct pseudotimes, rank 1 is assigned to the smallest value, rank 2 to the next smallest, and so on. Under any strictly increasing transformation, each pairwise comparison is unchanged. Therefore every cell retains its rank.

To see why, take any two cells \(i\) and \(j\). Strict increase means that whenever \(\tau_i < \tau_j\), \(g(\tau_i)<g(\tau_j)\); it also means the converse, since \(g(\tau_i)<g(\tau_j)\) cannot hold if \(\tau_i\geq\tau_j\). Thus the set of pairwise order relations is identical before and after transformation. Since ranks are determined by those pairwise relations, the ranks are identical as well. This justifies comparing rank patterns when two outputs may use different increasing scales. It does not justify comparing their raw differences as though those differences had a shared unit.

Assumptions, Diagnostics, and Limitations

Pseudotime interpretation is most defensible when the target process is plausibly represented by an ordered progression and the measured expression variation contains information about that process. The result also depends on how cells were sampled and measured. If batch effects or other technical differences align with the apparent progression, the ordering may reflect those differences. If cells from distinct biological populations are forced into one path, a single ordering can misrepresent the data. Uncertainty in cell identity can also matter when the path is interpreted using named cell types; annotation is evidence to consider, not a substitute for evaluating the inferred ordering.

Several checks help identify trouble without turning pseudotime into a biological clock:

• Check whether the proposed early and late ends agree with available biological anchors, such as known starting states or collection-time trends. Treat agreement as supporting evidence, not automatic proof.
• Inspect whether cells with similar labels or technical batches are segregated in a way that could dominate the ordering. A striking batch-aligned order raises concern that pseudotime reflects unwanted variation.
• Compare the cell ranks from reasonable analyses. Stable broad ordering supports a robust relative description; substantial rank changes signal sensitivity that should be reported.
• Look for ties, dense groups, or sparse regions. Tied or nearly tied values do not support a confidently resolved order between those cells, and a visually large numerical gap is not evidence of a long biological pause.
• Ask whether one path is an adequate summary. If distinct routes or populations are combined, a single total ordering may not capture their relationships. Interpret positions conditionally on the path being considered.

The main alternatives answer different questions. Actual collection times provide elapsed time since a defined event, but a cross-sectional time series does not necessarily track the same cells. An ordinal description or ranks emphasize relative ordering and avoid implying meaningful numerical spacing. A calibrated time model can aim to relate inferred position to measured time, but it needs time-linked data and explicit assumptions. Raw pseudotime coordinates are convenient for visualization and ordering, yet their units and gaps should not be given a stronger interpretation than the analysis supports.

Common Mistakes

• Wrong: “Cell X has pseudotime 0.8, so it is 0.8 hours into the process.” Correction: The value is an inferred coordinate unless a separate calibration establishes a time unit.
• Wrong: “The gap from 0.2 to 0.4 is twice as long biologically as the gap from 0.7 to 0.8.” Correction: Numerical gaps depend on the chosen scale; ordering alone does not establish durations or rates.
• Wrong: “Two methods agree because their values are numerically similar.” Correction: Compare the cells’ order and relevant anchors. Numerical similarity across arbitrary scales is not evidence of agreement.
• Wrong: “A known starting cell makes every pseudotime value a time measurement.” Correction: A root can fix orientation; it does not calibrate scale or prove the inferred path is correct.