Managing Redundant Genes in a Selected Feature Set

Why Redundancy Needs Careful Management

A feature-selection method can return several genes whose expression patterns across cells are nearly the same. Keeping all of them may give one co-expression pattern extra influence in a downstream analysis. Yet removing genes solely because they are correlated can erase distinct biological signals, including patterns associated with different cell populations. The goal is therefore not to make every pair of selected genes uncorrelated. It is to identify plausible duplication in the selected representation and reduce it without collapsing meaningful programs.

Here, a gene’s expression pattern is the vector of its expression values across a specified set of cells. Correlation describes the similarity of two such patterns on the chosen scale. It is a descriptive measure, not proof that genes have the same function, share a regulatory mechanism, or should always be represented by only one gene.

Definition: For a selected feature set, two genes are expression-redundant for a particular analysis when their measured patterns across the relevant cells are sufficiently similar that retaining both is judged to add little distinct information for that analysis. This is an operational judgment: it depends on the cell set, expression scale, similarity criterion, and intended downstream use. High positive correlation can flag a pair for review, but it does not by itself establish biological redundancy.

Measure Similarity on a Defensible Scale

Let \(x_{ig}\) denote the expression value for gene \(g\) in cell \(i\), after the chosen normalization and transformation. For two genes \(g\) and \(h\), Pearson’s sample correlation across \(n\) cells is

r_{gh} = \frac{\sum_{i=1}^{n}(x_{ig}-\bar{x}_g)(x_{ih}-\bar{x}_h)} {\sqrt{\sum_{i=1}^{n}(x_{ig}-\bar{x}_g)^2 \sum_{i=1}^{n}(x_{ih}-\bar{x}_h)^2}}.

It is defined when both genes have nonzero variation across the cells. Its value lies between \(-1\) and \(1\): values near \(1\) indicate similar linear patterns, values near \(-1\) indicate opposing linear patterns, and values near zero indicate little linear association. A high positive value is the most direct flag for duplicated co-expression; a high negative value can instead represent a meaningful opposing pattern and should not automatically be treated as a duplicate.

Correlation depends on which cells are included and how expression is represented. Use a common cell set and a scale appropriate to the analysis, such as suitably normalized and transformed expression or a chosen variance-stabilized representation. Raw counts can make library size and mean-dependent variability dominate the comparison. Conversely, transformations, residualization, or restricting the cells can change the relationship being measured. State the scale and cell population when reporting a correlation.

A cell-by-cell correlation is useful as a descriptive summary, but cells from one specimen are not independent biological replicates. Treating thousands of cells as independent replicates to obtain a tiny p-value would overstate the evidence. Redundancy management usually needs a transparent similarity rule and biological review, not a significance test at the cell level.

What a Correlation Threshold Does—and Does Not—Mean

A simple rule flags a pair when its correlation exceeds a chosen positive threshold, for example \(r_{gh} \geq 0.85\). This cutoff is a practical screening choice, not a universal biological boundary. The more pairs examined, the more likely it is that some exceed a threshold by chance or due to peculiarities of the data. The threshold should therefore be treated as a way to prioritize review, with sensitivity to nearby cutoffs considered when the result matters.

A useful justification for correlation-based screening comes from standardizing each gene’s vector across the same cells. Let \(z_g\) and \(z_h\) be vectors centered to mean zero and scaled to sample standard deviation one. Their squared Euclidean distance is

\lVert z_g-z_h\rVert^2 = \sum_{i=1}^{n}z_{ig}^2+\sum_{i=1}^{n}z_{ih}^2 -2\sum_{i=1}^{n}z_{ig}z_{ih}.

Because each standardized vector has sum of squares \(n-1\), and their cross-product is \((n-1)r_{gh}\), substitution gives

\lVert z_g-z_h\rVert^2=2(n-1)(1-r_{gh}).

Thus, for standardized expression patterns on the same cells, high positive correlation corresponds to small Euclidean distance. This equivalence explains why correlation can identify similar patterns; it does not establish that removing one gene preserves every biological interpretation.

Worked Example: Calculating a Pairwise Correlation

Suppose five hypothetical cells have normalized, transformed expression values \(x=(1,2,3,4,5)\) for gene A and \(y=(1,2,3,5,4)\) for gene B. The values are on the same scale and refer to the same cells. Their means are both \(3\), so the centered vectors are \((-2,-1,0,1,2)\) and \((-2,-1,0,2,1)\).

The cross-product sum is \(4+1+0+2+2=9\). Each centered sum of squares is \(4+1+0+1+4=10\). Therefore,

r_{AB}=\frac{9}{\sqrt{10\times 10}}=\frac{9}{10}=0.900.

If the screening rule is \(r \geq 0.85\), this pair is flagged for review. The result says that the observed patterns are strongly positively linearly associated in these five cells. It does not, by itself, say whether gene A or gene B should be removed, or whether either gene has a distinct functional interpretation.

Reduce Duplicates Without Collapsing Programs

A practical approach is to build a graph in which each selected gene is a node and an edge joins two genes whose positive correlation exceeds a stated threshold. This makes clusters of similar patterns visible. But an edge is a screening signal, not an instruction to delete a gene. Correlation is not transitive: A may be highly correlated with B and B with C even though A and C do not meet the threshold. Automatically reducing every connected component to one gene can therefore discard substantial variation.

Within a flagged group, consider which genes best represent the information needed downstream. A defensible rule might prioritize a gene with a stronger feature-selection score, clearer measurement, or a more useful interpretation, while retaining genes that represent distinct programs or cell populations. Record the rule and the retained representative. If a downstream task depends on a particular marker, that use may justify retaining it even when another selected gene is highly correlated.

Worked Example: Retaining a Representative, Not Deleting a Program

Consider four hypothetical selected genes measured in the same four cells on a transformed scale. Gene A has values \((1,2,3,4)\), gene B has \((2,4,6,8)\), gene C has \((4,3,2,1)\), and gene D has \((1,1,2,2)\). Suppose A and B have feature scores 2.1 and 2.8, respectively. Use \(r \geq 0.90\) to flag positive similarity, and suppose the analysis aims to represent co-expression programs rather than preserve every gene as an individual marker.

For A and B, the means are 2.5 and 5. The centered vectors are \((-1.5,-0.5,0.5,1.5)\) and \((-3,-1,1,3)\). The cross-product sum is \(4.5+0.5+0.5+4.5=10\); the sums of squares are \(5\) and \(20\). Hence

r_{AB}=\frac{10}{\sqrt{5\times20}}=1.000.

The pair is perfectly positively correlated in these hypothetical data. If the stated policy is to retain the higher-scoring representative from a strongly co-expressed pair, retain B and mark A as redundant for this particular representation. This decision follows the declared rule; it does not imply that A has no biological role.

Gene C has the reverse pattern to A, so its correlation with A is \(-1\), not a high positive correlation. Do not remove C under this positive-correlation rule: opposing expression may encode a distinct pattern. Gene D also should not be removed simply because it appears in the same selected set; its relationship to the others requires assessment. The outcome is a focused reduction of the A–B duplication, not an automatic removal of every gene associated with the same broad biological theme.

Check for Cell-Composition and Other Distortions

A pooled correlation can reflect differences between cell populations rather than co-expression within either population. A gene pair may look redundant across a mixed set of cells because both are higher in one population, even when their within-population patterns differ. Conversely, a pooled summary can hide a relationship that is clear within particular populations. Correlations should therefore be interpreted in light of cell composition and the question being asked. Where appropriate, inspect patterns within relevant groups as well as across the pooled cells; do not interpret a within-group estimate as a substitute for the pooled one without stating the change in target.

Worked Example: A Pooled Correlation Can Mislead

Suppose two hypothetical genes are measured in six cells from two equally sized populations. In population 1, gene X is \((1,2,3)\) and gene Y is \((3,2,1)\). In population 2, X is \((7,8,9)\) and Y is \((9,8,7)\). Within each population, the vectors are exact reversals, so the within-population correlation is \(-1\).

Pooling the cells gives \(x=(1,2,3,7,8,9)\) and \(y=(3,2,1,9,8,7)\). Both means are \(5\). The centered vectors are \((-4,-3,-2,2,3,4)\) and \((-2,-3,-4,4,3,2)\). Their cross-product sum is \(8+9+8+8+9+8=50\); each sum of squares is \(58\). Thus,

r_{\text{pooled}}=\frac{50}{\sqrt{58\times58}} =\frac{50}{58}\approx 0.862.

A pooled threshold of \(0.85\) would flag these genes as positively correlated, despite their negative within-population relationships. In this constructed example, separation between the two populations drives the pooled association. The appropriate interpretation is not “these genes are redundant,” but “their pooled patterns are strongly associated, and the association changes within populations.” Whether to represent the population contrast, the within-population behavior, or both depends on the analysis.

Diagnostics and Alternatives

Before reducing a set, check that the correlations use the intended cell population and expression representation, and that each gene has adequate variation for the statistic to be meaningful. Inspect scatterplots or expression summaries for flagged pairs: a high Pearson correlation can be driven by a small number of cells or by separation between groups. Also inspect the genes’ identities and known or hypothesized roles, while treating such knowledge as context rather than proof. A few practical alternatives answer different questions:

Approach	Useful when	Main caution
Pairwise Pearson correlation	Similarity is expected to be approximately linear on the chosen expression scale.	Sensitive to cell composition, extreme values, and the selected scale.
Rank correlation	A monotonic relationship matters more than a linear one.	Tied values, including many zeros, can limit its resolution; it does not solve composition confounding.
Correlation graph or clustering	Many pairs need to be organized into groups of similar patterns.	Results depend on the similarity rule and grouping procedure; a cluster is not automatically one biological program.
Retaining all selected genes	Gene-level interpretation or marker coverage is important.	Highly similar features can give a shared pattern extra weight in some downstream analyses.

These approaches are not interchangeable. Rank correlation changes the type of association summarized; clustering organizes pairwise similarities; retaining all features avoids an arbitrary deletion but does not remove duplicated representation. None alone establishes biological redundancy. The selection should match the downstream purpose, and correlation reduction should not be used as a substitute for addressing normalization problems or cell-population structure.

Common Mistakes and Practical Limits

“The correlation is high, so the genes are duplicates.” High correlation describes similar patterns in the specified cells and scale. It does not establish common function or make either gene dispensable.
“I will remove every gene connected to a high-correlation pair.” A pairwise edge is not a deletion rule, and correlation is not transitive. Review groups and use an explicit representative-selection policy.
“Negative correlation is just another kind of redundancy.” Opposing patterns can encode meaningful biological contrasts. Do not treat a large absolute correlation as automatic grounds for removal.
“A very small cell-level p-value proves the genes are redundant.” Cells from the same biological sample are not independent replicates, and statistical association is not proof of interchangeability. Use correlation here as a descriptive screen.
“One pooled correlation answers every question.” Cell composition can create or obscure associations. Check whether the pooled cell set matches the intended interpretation.

Redundancy management is necessarily limited by the data and the analysis goal. A single correlation cannot capture every nonlinear or population-specific relationship, and a reduced set may be less suitable for a different downstream task. Document the scale, cells, criterion, and retention rule so the decision can be interpreted and revisited.

Check Your Understanding

Answer each question using the definitions and reasoning above. Brief solutions follow.

For two genes with \(r=0.92\) on a stated transformed scale, what can be concluded, and what cannot be concluded?
Why should a high negative correlation not automatically trigger removal under a positive co-expression rule?
A pair exceeds a correlation threshold in pooled cells but has weak correlations within two cell populations. Name one plausible explanation and one diagnostic step.
Why is it risky to keep one gene from every connected component of a thresholded correlation graph?

Solutions. (1) Their patterns are strongly positively linearly associated across the specified cells on that scale; this alone does not show shared function or prove either gene can be removed. (2) The genes may encode opposing, biologically distinct patterns, and a positive co-expression rule targets similar rather than opposite expression. (3) Population composition may drive the pooled association; inspect the expression patterns and correlations within the populations as well as in the pooled set. (4) Correlation is not transitive, so members linked through intermediate genes need not all be highly similar to one another; reducing an entire connected component to one gene can discard distinct patterns.