Why a Principal Curve Needs a Direction
A principal curve describes a path through a cell representation, and projection onto that path gives each cell a location along it. But the geometry of an open path does not say which end is biologically earlier. If the same fitted curve is traversed in the opposite direction, its shape and cell-to-curve distances are unchanged. The coordinates called “early” and “late,” however, are exchanged.
A biologically specified starting cell or state supplies the missing directional information. It might be a cell measured at an experimentally defined initial time, or a population known to precede the transition being studied. Orienting a curve means choosing which end corresponds to that starting point and assigning pseudotime to increase away from it. This is a decision about the interpretation of a fitted path, not a new curve-fitting step.
The definition assumes a single, nonbranching, open curve with a meaningful pair of ends, and a fitted geometry that is being held fixed while direction is assigned. It also assumes the biological root corresponds to one end, or at least lies close enough to one end that choosing that end is justified. Under these conditions, the root resolves a directional ambiguity; it does not establish that the curve is a correct model of the cells’ biology.
Turning a Root into an Arc-Length Coordinate
With the parameterization above, one end is represented by \(s=0\) and the other by \(s=L\). If the root is at the \(s=0\) end, the oriented pseudotime is \(\tau_i=s_i\). If the root is at the \(s=L\) end, pseudotime is \(\tau_i=L-s_i\). Thus, in either case, the root end has pseudotime zero and values increase along the curve away from that end. Pseudotime has the units of the distance used to parameterize the curve; when the embedding has no physical units, it is an embedding-distance scale rather than elapsed time.
The formula presumes the fitted curve has already been given an arbitrary coordinate direction. That initial choice is only bookkeeping: reversing the coordinate changes \(s_i\) to \(L-s_i\), but not the curve itself. The root determines which of the two equivalent coordinate directions should receive the interpretation “increasing pseudotime.”
For a root cell that projects a short distance from an end, it is usually clearer to identify the corresponding curve end as the origin than to force that one cell to have exactly zero pseudotime. Projection error and variation within a root population can place biologically early cells slightly along the curve. In that situation, nearby root cells should have small nonnegative pseudotimes, rather than being assigned negative values merely to make one noisy projection equal zero.
Worked Example: Reversing the Labels at the Opposite Root
Suppose a fitted curve is a two-segment path through \((0,0)\), \((3,0)\), and \((3,4)\), with distances measured in embedding units. Its length is \(L=3+4=7\). Four hypothetical cells project to arc-length coordinates \(s_A=0.5\), \(s_B=2\), \(s_C=4\), and \(s_D=6.5\).
If the biological root is at the first endpoint, \(s=0\), then \(\tau=s\). The pseudotimes are \(0.5\), \(2\), \(4\), and \(6.5\) embedding units for cells A through D. Their order from early to late is A, B, C, D.
If instead the known root is at the other endpoint, \(s=7\), then \(\tau=7-s\). Substitution gives \(\tau_A=7-0.5=6.5\), \(\tau_B=7-2=5\), \(\tau_C=7-4=3\), and \(\tau_D=7-6.5=0.5\) embedding units. The order from early to late is now D, C, B, A.
The curve length and the projected locations have not changed. Only the biological assignment of direction has changed. These pseudotime values are relative path coordinates, not measured durations in hours or days.
Choosing the Root End from a Known Cell or State
If a known root is a single cell, project it onto the fitted curve and compare its projected coordinate with the two ends. A projection close to \(s=0\) supports orienting pseudotime as \(s_i\); a projection close to \(s=L\) supports \(L-s_i\). This comparison is meaningful only if the chosen embedding, distance, and curve are the ones used to calculate the other cells’ projections.
A root state often contains multiple cells rather than one unique anchor. Project those cells and examine the distribution of their coordinates. A compact group near one end can provide stronger evidence for the root end than an isolated cell. A simple summary, such as the median projected coordinate, can help describe the group’s location, but should not conceal substantial spread: a group split between the ends or spread across much of the curve does not specify a clear root end.
The biological information used to specify a root should be independent of the arbitrary direction assigned during curve fitting. For example, an experimentally defined starting time can support a root assignment. Merely calling the end with lower expression of a selected gene “early” is not, by itself, independent evidence that the cells are biologically earlier. Expression trends can help describe a rooted trajectory, but using a trend to choose the root and then treating the same trend as confirmation risks circular reasoning.
Worked Example: Orienting from a Root Population
Consider a hypothetical fitted curve of length \(L=10\) embedding units. Three cells from a biologically specified initial state project at \(s=0.3\), \(0.5\), and \(0.7\). Their median coordinate is \(0.5\). The average distance of these projections from the \(s=0\) endpoint is \((0.3+0.5+0.7)/3=0.5\) units.
Their distances from the other endpoint, \(s=10\), are \(9.7\), \(9.5\), and \(9.3\) units, with average \((9.7+9.5+9.3)/3=9.5\) units. The root population therefore lies much closer to the \(s=0\) end in this example. If the biological root assignment is credible and the projections are on the same fitted curve, orient the pseudotime as \(\tau_i=s_i\).
The root cells then have pseudotimes \(0.3\), \(0.5\), and \(0.7\) embedding units. They are close to the root end but need not have pseudotime exactly zero. A cell projecting to \(s=4.2\) has pseudotime \(4.2\) embedding units. This calculation identifies a direction; it does not show that the curve’s endpoints or intermediate geometry are biologically correct.
Why Reversal Changes Interpretation but Not Geometry
The invariance to reversal follows directly from the arc-length parameterization. Define a reversed curve by \(\hat{\gamma}_{\mathrm{rev}}(s)=\hat{\gamma}(L-s)\). It visits the same points as the original curve in the opposite order. For any cell \(i\), its distance to a point on the reversed curve is its distance to the same point on the original curve. Consequently, the set of closest curve points and the curve’s geometric fit are not improved or worsened merely by reversing direction.
The coordinate of the corresponding projection is reversed. If its original coordinate is \(s_i\), its reversed coordinate is \(s_i^{\mathrm{rev}}=L-s_i\). When the root is at the original \(s=L\) end, the root-oriented pseudotime is therefore \(L-s_i\). The relative order of cells is reversed, even though the projected points are identical.
This invariance also clarifies what orientation cannot do. If iterative curve fitting has settled in a poor local solution, selecting a root end does not move the curve into a better solution. The earlier material on initialization and local minima concerns the fitted geometry; root orientation assigns a biological direction to a geometry that has already been obtained.
Worked Example: An Expression Trend Changes Sign on Reversal
Suppose four hypothetical cells lie at coordinates \(s=0,1,2,3\) on a curve of length \(L=3\). A measured gene’s expression values are \(2,5,8,11\) arbitrary expression units, respectively. Along the arbitrary \(s\) coordinate, these values follow \(y=2+3s\).
Now suppose the known root is at \(s=3\). The root-oriented pseudotimes are \(\tau=3-s\), so the four cells have pseudotimes \(3,2,1,0\), paired with expression values \(2,5,8,11\). Since \(s=3-\tau\), substitution into the original expression relation gives \(y=2+3(3-\tau)=11-3\tau\).
The expression slope is therefore \(+3\) expression units per embedding unit with respect to \(s\), but \(-3\) expression units per embedding unit with respect to root-oriented \(\tau\). No expression values or cell positions changed. The sign changed because the coordinate direction changed. A trend along pseudotime is an association with position on the fitted curve; by itself, it does not establish a causal effect or a rate of change over clock time.
When a Root Does Not Define One Directed Ordering
A root in the interior of an open curve does not, in general, determine a single nonnegative pseudotime that increases away from the root across the whole curve. If the root coordinate is \(s_r\), cells on one side have \(s_i<s_r\) and cells on the other have \(s_i>s_r\). The signed coordinate \(s_i-s_r\) distinguishes the sides, but increases in only one direction along the curve. The unsigned distance \(\lvert s_i-s_r\rvert\) increases away from the root on both sides, but maps positions on opposite sides to the same value and cannot order those cells relative to each other.
Thus an interior root may be informative about a local starting region, but it does not supply the ordinary one-way orientation of a single open trajectory. Do not silently treat either side as later than the other. A branched biological process may require a representation that preserves separate paths; choosing a principal-curve direction cannot resolve that structural issue.
A root can also be ambiguous when its cells project far from either endpoint, have widely dispersed coordinates, or are comparably close to both ends. In those cases, the appropriate conclusion is that the root evidence does not clearly orient this fitted curve. Re-check the biological assignment, projections, and curve representation rather than choosing the direction that produces a preferred expression pattern.
Use the same embedding, distance, and projected arc-length coordinates used to fit and describe the curve.
Project the specified root cell or root-state cells and inspect their coordinates relative to both ends.
Use the endpoint closest to a credible root as pseudotime zero; if the evidence is interior, dispersed, or ambiguous, report that a single orientation is not established.
Use \(s_i\) or \(L-s_i\) as appropriate, state the distance units, and interpret the values as relative positions along the curve.
Comparison, Diagnostics, and Limitations
Root-based orientation is one way to label direction, not a test of whether the fitted trajectory is biologically valid. It differs from leaving the curve unoriented: an unoriented curve supports statements about proximity and relative location, but not about which end is earlier. It also differs from choosing a direction solely from gene expression. A known experimental or biological starting state supplies an external anchor, while expression-based direction depends on assumptions about which genes should rise or fall.
| Approach | What it contributes | Main limitation |
|---|---|---|
| Known root cell | Uses a specific biological anchor to select an end. | Atypical or noisy projection of one cell may be unrepresentative. |
| Known root state | Uses a population of cells to assess which end is supported. | Diffuse or mixed projections may not identify one end. |
| Expression-based direction | Uses a selected expression pattern to suggest increasing or decreasing direction. | Can be circular if the same pattern is used to define and validate the direction. |
| Unoriented curve | Preserves the fitted geometry without making an unsupported early-to-late claim. | Does not label a starting end or directed progression. |
Several checks help prevent an arbitrary choice from being mistaken for biological evidence. Confirm that the root cells are represented in the same analysis space and are projected onto the same curve as the other cells. Check whether their coordinates consistently favor one end rather than relying on one extreme cell. Confirm that increasing pseudotime actually moves away from the selected root end. Finally, report that the values are relative curve positions; a path length is not elapsed time unless an independent calibration supplies a time scale.
A principal curve can also be an inadequate representation for the biological process, and root orientation cannot correct that mismatch. In particular, a single open path cannot encode distinct outgoing branches or a cycle as one simple increasing order. These are limitations of the representation, not problems solved by reversing the parameterization. Likewise, a root chosen from incomplete or uncertain biological evidence gives a conditional interpretation: the oriented pseudotime follows if that root assignment is appropriate.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A curve has length \(8\) embedding units. A cell projects to \(s=2.5\), and the known root is at \(s=8\). What is its root-oriented pseudotime, and what are its units?
- Why does reversing the direction of a fitted principal curve leave the distances from cells to the curve unchanged?
- A root-state population projects at coordinates \(3.8\), \(4.0\), and \(4.2\) on a curve of length \(8\). Does this evidence clearly identify an endpoint as the root? Explain using distances.
- For a curve of length \(5\), an expression trend is \(y=1+2s\), and the root is at \(s=5\). Express the trend in root-oriented pseudotime \(\tau\).
- What is the difficulty with defining ordinary one-way pseudotime when the root lies in the interior of an open curve?
Solutions. (1) Since the root is at \(s=L\), \(\tau= L-s=8-2.5=5.5\) embedding units. (2) Reversal visits the same geometric points in the opposite order; each cell’s closest point is still the same point in space, so its minimum distance is unchanged. (3) The average distance to \(s=0\) is \((3.8+4.0+4.2)/3=4.0\) units. The average distance to \(s=8\) is \((4.2+4.0+3.8)/3=4.0\) units. The population is centered and does not favor either endpoint, so it does not clearly identify one. (4) Since \(\tau=5-s\), \(s=5-\tau\), and \(y=1+2(5-\tau)=11-2\tau\). The slope is \(-2\) expression units per embedding unit of root-oriented pseudotime. (5) Cells on opposite sides of an interior root extend in opposite curve directions. Signed distance distinguishes the sides but does not increase away from the root on both; unsigned distance increases on both sides but does not preserve a unique order between them.