Why a Plausible Anchor Can Be Wrong
Anchor-based integration uses pairs of cells judged to represent corresponding states across datasets to estimate how one dataset should be adjusted relative to another. This is useful when a technical shift separates genuinely shared states. But geometric proximity is not proof of biological correspondence. If a state occurs only in one condition, or if a transitional state lies between familiar populations, an algorithm may pair it with a nearby but biologically different state. The resulting correction can make the datasets look more aligned while erasing the distinction the analysis needs to preserve.
The central question is therefore not merely whether anchors can be found, but whether their partners represent the same underlying state. This matters especially when condition and batch are entangled: a condition-specific population in one batch may have no counterpart in the reference, even though some reference cells occupy nearby positions in a latent space.
What an Anchor Correction Assumes
Let \(z_{Q,i}\) be the coordinate of cell \(i\) in a query dataset and \(z_{R,j}\) the coordinate of a cell in a reference dataset, using the same \(p\)-dimensional representation. For an anchor pair \(\ell=(i_\ell,j_\ell)\), define the correction vector for the query relative to the reference as \(d_\ell=z_{R,j_\ell}-z_{Q,i_\ell}\). If cell \(i\) receives weights \(w_{i\ell}\) from its nearby anchors, with \(w_{i\ell}\geq 0\) and \(\sum_\ell w_{i\ell}=1\), a weighted correction has the form \(z^{\mathrm{corr}}_{Q,i}=z_{Q,i}+\sum_\ell w_{i\ell}d_\ell\). This expression describes the direction of the adjustment; it does not establish that the anchor pairs are biologically valid.
A useful conceptual model for a cell in state \(s\), batch \(b\), and condition \(c\) is \(z_{b,s,c}=\mu_{s,c}+\delta_{b,s,c}+\epsilon\). Here \(\mu_{s,c}\) is the biological location of that state under that condition, \(\delta_{b,s,c}\) is a batch-related displacement, and \(\epsilon\) is cell-level variation. For a genuinely shared state in comparable conditions, a cross-batch difference may mainly reflect the batch displacement. For a false pair between different states, its difference also contains a biological-state term. An anchor correction cannot know which part of the difference is technical merely from the vector itself.
The key assumptions are consequently substantive as well as geometric: paired cells represent the same biological state; the datasets have adequate shared-state overlap to estimate the correction; and local correction vectors are relevant to the cells receiving them. If condition and batch are perfectly confounded, the observed coordinates alone generally cannot identify which differences are technical and which are biological. More anchors do not resolve that lack of information.
How False Correspondence Enters the Correction
The source of the problem can be seen by separating two kinds of anchor difference. Suppose a query cell is in state \(t\), but is paired to a reference cell in state \(s\). Ignoring random variation, their difference is \(z_{R,s}-z_{Q,t}=(\mu_s-\mu_t)+(\delta_R-\delta_Q)\). The first term is a biological difference between states; the second is a batch difference. The correction treats the combined difference as a direction for alignment. Unless \(s=t\), it can therefore use the biological distinction as if it were batch displacement.
This is not fixed simply by averaging several anchors. Averaging reduces the influence of noisy vectors when those vectors estimate the same underlying shift. If the vectors instead point from a condition-specific state toward different reference populations, their average may still move that state toward an inappropriate location. The following derivation makes the risk explicit.
Thus, in the simplest case, correction maps the query cell exactly to its false partner. With several anchors, the result is a weighted compromise among their partners. That compromise need not correspond to any real biological state.
Worked Example: A Condition-Specific Population With No Match
Suppose a reference contains cells from state B near coordinates \((4.0,0.0)\) and \((4.2,0.1)\). A query contains a condition-specific state T near \((3.8,0.2)\), with no true T population in the reference. Assume, hypothetically, that the anchor procedure retains the two query-to-reference pairs and assigns each weight \(0.5\) for this query cell.
The two anchor correction vectors are \(d_1=(4.0,0.0)-(3.8,0.2)=(0.2,-0.2)\) and \(d_2=(4.2,0.1)-(3.8,0.2)=(0.4,-0.1)\). Their weighted average is \(0.5(0.2,-0.2)+0.5(0.4,-0.1)=(0.3,-0.15)\).
The corrected coordinate is therefore \((3.8,0.2)+(0.3,-0.15)=(4.1,0.05)\). This lies between the two reference B cells. The arithmetic is internally consistent, but the result does not show that T and B are the same state. Given the assumed absence of a reference T population, the apparent alignment is evidence of a potentially false match, not successful recovery of a shared state.
Misleading Neighborhoods Along a Transition
A transitional population creates a related failure. A cell along a biological continuum may lie between two established populations in the chosen representation. If cells at either end are well represented in the reference, nearby anchors can point in opposing directions. The resulting correction depends on their weights. It may pull the transitional cells toward one endpoint even when the transition itself is condition-specific or biologically meaningful.
Worked Example: A Transitional Cell Between Two Reference States
In a one-dimensional representation, let reference state A be at coordinate \(0\), reference state B at coordinate \(4\), and a query transitional cell U at coordinate \(2\). Suppose U receives an anchor to A with weight \(0.75\) and an anchor to B with weight \(0.25\). These weights sum to one, as required for the weighted correction.
The correction vectors are \(d_A=0-2=-2\) and \(d_B=4-2=2\). Their weighted average is \(0.75(-2)+0.25(2)=-1.0\). The corrected coordinate is \(2+(-1)=1\).
The cell has moved one coordinate unit toward A. If U represents a real transitional state centered at 2, this adjustment distorts its position; it does not establish that U is a technical version of A. With reversed weights, \(0.25\) for A and \(0.75\) for B, the average vector would be \(0.25(-2)+0.75(2)=1\), moving U to 3 instead. The direction of distortion depends on the local anchor composition.
A stable-looking correction can still be wrong: if anchors on both sides balance, their weighted vectors may nearly cancel, leaving the cell near its original location. Conversely, a small change in neighborhood composition may produce a large directional change. For transitional states, inspect the partners and vector directions, not only the final corrected coordinates.
Diagnosing Condition-Specific Anchor Failures
The earlier material on diagnosing anchor coverage and directionality provides a starting point. For condition-specific states, the diagnosis must also ask whether anchor partners are biologically comparable and whether the correction preserves meaningful condition structure. Useful checks include:
- Inspect partner identity. Where labels or marker evidence are available, tabulate the state and condition of both cells in each anchor pair. Concentration of partners from a different state is a warning, even when distances are small.
- Check one-sided coverage. Ask whether a state in one dataset has a plausible counterpart in the other. Many query cells assigned to the same reference state, with no reciprocal evidence of that population, may indicate an unmatched population rather than broad shared-state support.
- Examine correction-vector coherence. Within a proposed state, compare vector directions and magnitudes. Strongly opposing vectors, abrupt direction changes, or unusually large shifts can indicate that the neighborhood combines different biological states.
- Compare before and after integration. Verify that known condition-specific markers, trajectories, or state separations have not disappeared merely because the corrected coordinates overlap.
- Vary the local evidence. If modest changes in neighborhood size or representation produce different partner states or correction directions, the claimed correspondence is fragile.
These checks are not proofs that an anchor is valid. Labels may be incomplete, marker expression may vary with condition, and latent-space geometry depends on representation choices. They make the assumptions visible and help distinguish supported alignment from forced alignment.
Worked Example: Separating a Shared Batch Shift From a False Match
Suppose a shared state S has reference coordinate \(0\) and query coordinate \(1\), so the observed query-reference difference for this shared state is \(1\). A query condition-specific state T is at coordinate \(4\), while the reference contains no T. Assume a false anchor pairs T with reference S at \(0\).
The false anchor vector is \(0-4=-4\). Applying it gives \(4+(-4)=0\), placing T on top of reference S and erasing their four-unit separation in the query representation.
For the genuinely shared state S, the observed difference is \(0-1=-1\). If, as a simplifying hypothetical assumption, that one-unit displacement represents a common technical shift, applying \(-1\) to the query T coordinate would instead give \(4-1=3\). The separation between corrected T and reference S would remain three units. This comparison shows why a shared-state correction and a false cross-state anchor imply different adjustments. It does not prove that the shift is common: that assumption needs support from shared-state evidence.
The conclusion is limited to the stated coordinates and assumptions. If the condition effect itself changes the location of S, or if the batch displacement varies by state, the shared-state difference may not estimate the technical shift for T.
Choosing a Response When Evidence Is Weak
When anchor validity is uncertain, the appropriate response is not automatically to force more alignment. The alternatives have different costs: leaving datasets uncorrected can retain technical separation, while aggressive correction can erase biology. The choice should follow the scientific target and the evidence for overlap.
| Approach | When it can help | Main limitation |
|---|---|---|
| Correct using anchors across all nearby cells | Shared states are well supported and local batch shifts are plausible. | False partners can transfer biological differences into correction. |
| Restrict correction to supported shared states | Some populations have credible counterparts while others do not. | Unmatched cells remain unaligned; this is not necessarily a defect. |
| Withhold integration for the uncertain population | There is insufficient evidence to separate condition from batch effects. | Technical variation remains and must be acknowledged in downstream comparisons. |
| Analyze conditions separately or preserve condition labels | The scientific question concerns condition-specific states or transitions. | Direct cross-condition comparisons may be less precise or less convenient. |
Reference-based mapping can help locate query cells relative to a reference, but a mapped location is not proof that a query population has a true biological counterpart there. In particular, an out-of-reference state should not be interpreted as a known reference state merely because it receives a nearby assignment. Whatever approach is used, report which cells were aligned, which were not, and what evidence supports the shared-state assumption.
Common Pitfall and Practical Limits
Wrong reasoning: “The integration produced strong overlap, so the two populations must be the same cell type.” This reverses the logic. Overlap is partly the result of the correction procedure, so it cannot independently validate the correspondence that generated that correction.
Correction: Evaluate biological correspondence using evidence that is not just the corrected geometry: state markers, condition context, partner identities, and whether the state is plausibly present in both datasets. If no counterpart is supported, treat that population as unmatched rather than interpreting forced overlap as discovery.
There are limits to what diagnostics can establish. If every cell in one condition was processed in one batch and every cell in another condition in a different batch, batch and condition effects may be inseparable from these data alone. No anchor weighting scheme can recover information that the design does not provide. A conservative result may retain apparent batch structure while preserving condition-specific biology; that can be preferable to a visually cleaner embedding with unsupported biological equivalences.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A query-only population is close to reference cells of one known state. Why does proximity alone not justify treating the populations as shared?
- A query cell at coordinate \(5\) is paired to a reference cell at coordinate \(2\) with weight \(1\). What correction is applied, and what biological risk arises if the pair is across states?
- Two anchors for a transitional query cell have vectors \(-2\) and \(2\), with weights \(0.4\) and \(0.6\). Calculate the average correction vector. What does the result say, and what does it not say?
- Name two diagnostics that can reveal a misleading shared neighborhood.
- Why can perfect confounding of condition and batch prevent a confident distinction between technical and biological differences?
Solutions. (1) Proximity describes the selected representation, not biological identity; distinct states can be nearby, and the population may have no reference counterpart. (2) The vector is \(2-5=-3\), so the corrected query coordinate is \(5-3=2\). If the cells differ biologically, this maps the query state onto the reference state and may erase that difference. (3) The average is \(0.4(-2)+0.6(2)=-0.8+1.2=0.4\). The correction points in the positive direction by \(0.4\) units; it does not establish that either anchor is a valid match or that the movement is biologically appropriate. (4) Examples include inspecting anchor-partner state labels and checking whether correction vectors are coherent within the proposed state. One-sided coverage and preservation of condition-specific markers are also useful. (5) When each condition occurs only in a different batch, observed differences combine condition and batch effects; without additional information or design overlap, those components are not separately identified.