Leading-Edge Genes in GSEA Results

Why the Leading Edge Deserves a Closer Look

A GSEA enrichment score compresses a ranked list and a gene set into one number: the largest positive or negative departure of a running sum from zero. That score tells us which direction and how far the running sum departs, but not which pathway genes occur on the relevant side of the extremum. The leading-edge subset supplies that positional summary. It can help identify genes shared by several enriched pathways, while its definition also imposes limits on what the subset can establish.

The interpretation depends on the ranked list, the gene-set membership, the hit-weighting rule, and the location of the selected extremum. The definition is descriptive: it does not, by itself, provide a new significance test or show that the listed genes cause a phenotype.

Definition From the Running Sum

Let \(g_1,\ldots,g_N\) be the genes in a ranked list, in their chosen order, and let \(S\) be a gene set represented in that list. Write \(m=|S|\) for the number of set members in the list. The construction below assumes \(0<m<N\), that gene identifiers have been mapped consistently, and that the hit-weight total is positive. For a weighted analysis with exponent \(p>0\), a hit at position \(j\) has weight \(w_j=|r_j|^p\), where \(r_j\) is its ranking statistic; let \(W=\sum_{g_j\in S}w_j\). For the unweighted case \(p=0\), define every hit weight to be 1, so \(W=m\).

Starting at zero, the running sum \(R(k)\) adds a normalized weight at a pathway hit and subtracts an equal step for each miss. A positive-weight hit raises the sum; a zero-weight hit leaves it unchanged; a miss lowers it. In particular, when \(p>0\), a hit with \(r_j=0\) has weight zero and contributes no increase. The update is:

R(0)=0,\qquad R(k)=R(k-1)+ \begin{cases} w_k/W, & g_k\in S,\\ -1/(N-m), & g_k\notin S. \end{cases}

The enrichment score (ES) is the signed extremum with the larger absolute departure from zero: the maximum of \(R(k)\) if the positive extremum is at least as large in absolute value as the negative one, and otherwise the minimum. The position \(k^*\) is where that selected extremum occurs. When tied positions could lead to different subsets, the tie-handling rule should be documented; software may make a particular implementation choice.

Definition: For a positive ES at position \(k^*\), the leading-edge subset is the members of \(S\) at or before that position: \(\{g_j:g_j\in S,\ j\leq k^*\}\). For a negative ES, it is the members of \(S\) at or after the minimum position: \(\{g_j:g_j\in S,\ j\geq k^*\}\). These are the pathway genes on the side of the ranked list specified by the selected extremum. A zero-weight hit can meet this positional definition, even though it contributes no hit increment.

This definition follows the conventional directional interpretation of GSEA leading edge. For positive enrichment, the relevant pathway members accumulate toward the top of the ranked list before the peak. For negative enrichment, the subset consists of pathway members in the lower-ranked tail from the trough onward. In the negative case, it is important not to say that each listed hit itself drove the running sum down: hits raise or leave the sum unchanged, whereas misses lower it.

Worked Calculations

Worked Example: Finding a Positive Leading Edge

Suppose a hypothetical ranked list has \(N=8\) genes, and the pathway contains the genes at positions 1, 3, 6, and 8. Use a weighted analysis with \(p=1\), with absolute ranking-statistic values 5, 3, 0.5, and 2 at those positions, respectively. All other positions are misses. The hit-weight total is \(W=5+3+0.5+2=10.5\); there are \(N-m=4\) misses, each contributing \(-1/4=-0.25\). The assumptions \(0<m<N\) and \(W>0\) hold.

The hit increments are \(5/10.5=0.4762\), \(3/10.5=0.2857\), \(0.5/10.5=0.04762\), and \(2/10.5=0.1905\). Applying the update at every position gives:

\begin{aligned} R(1)&=0+0.4762=0.4762,\\ R(2)&=0.4762-0.25=0.2262,\\ R(3)&=0.2262+0.2857=0.5119,\\ R(4)&=0.5119-0.25=0.2619,\\ R(5)&=0.2619-0.25=0.01190,\\ R(6)&=0.01190+0.04762=0.05952,\\ R(7)&=0.05952-0.25=-0.1905,\\ R(8)&=-0.1905+0.1905=0. \end{aligned}

The maximum is \(0.5119\) at position 3, which is larger in absolute value than the minimum \(-0.1905\). Thus the ES is positive, and the leading edge consists of the pathway genes at positions 1 and 3. The hits at positions 6 and 8 are not in this leading edge because they occur after the selected peak.

Worked Example: Interpreting a Negative Leading Edge

In a second hypothetical analysis, let \(N=8\), use unweighted GSEA (\(p=0\)), and suppose the pathway genes are at positions 5, 7, and 8. Then \(m=3\), each hit adds \(1/3\), and each of the \(N-m=5\) misses subtracts \(1/5=0.2\). This is defined because there are both hits and misses, and the unweighted hit total is \(W=3\).

The running sum is \(-0.2,-0.4,-0.6,-0.8\) through the first four misses. The hit at position 5 gives \(R(5)=-0.8+1/3=-0.4667\). The miss at position 6 gives \(R(6)=-0.6667\); the hit at position 7 gives \(R(7)=-0.3333\); and the hit at position 8 returns the sum to zero. The minimum is \(-0.8\) at position 4, so the ES is negative. By the negative-direction definition, the leading edge comprises the pathway genes at positions 5, 7, and 8.

Here the trough occurs before any pathway hit. The misses in the upper part of the list make the running sum more negative; the listed leading-edge hits are the pathway members in the tail on the trough side. This example shows why “leading-edge genes caused the negative score” is not a careful interpretation.

Worked Example: Comparing Leading Edges Across Pathways

Use the hypothetical eight-gene list from the first example, with ranking-statistic values \(5,4,3,2,1,0.5,-1,-2\) in order. Pathway A has members at positions 1, 3, 6, and 8, and pathway B has members at positions 1, 2, and 4. With \(p=1\), pathway A has the running sum already calculated and leading edge \(\{g_1,g_3\}\).

For pathway B, the hit weights are 5, 4, and 2, so \(W=11\). There are five misses, each with decrement \(-1/5=-0.2\). The running sum by position is:

\begin{aligned} R(1)&=5/11=0.4545,\\ R(2)&=0.4545+4/11=0.8182,\\ R(3)&=0.8182-0.2=0.6182,\\ R(4)&=0.6182+2/11=0.8000,\\ R(5)&=0.8000-0.2=0.6000,\\ R(6)&=0.6000-0.2=0.4000,\\ R(7)&=0.4000-0.2=0.2000,\qquad R(8)=0. \end{aligned}

The maximum is \(0.8182\) at position 2, so pathway B’s leading edge is \(\{g_1,g_2\}\). The two leading edges share one gene, \(g_1\); their union has three genes. Their Jaccard overlap, the intersection size divided by the union size, is therefore \(1/3=0.3333\). This overlap is a compact descriptive comparison, not a test of whether the pathways share a biological mechanism.

Why the Running Sum Returns to Zero

The end-point property is a useful check on hand calculations. Across the full ranked list, the normalized hit increments sum to 1 because their numerators sum to \(W\), and the miss decrements sum to \(-1\) because there are \(N-m\) misses. Thus the endpoint is zero. This argument requires a defined hit total and at least one miss:

\sum_{g_j\in S}\frac{w_j}{W} +\sum_{g_j\notin S}\left(-\frac{1}{N-m}\right) =\frac{W}{W}-\frac{N-m}{N-m}=0.

If \(p>0\) and every pathway hit has \(r_j=0\), then \(W=0\) and the weighted hit increments are undefined. The analysis cannot use this weighted construction without changing the weighting specification. Individual zero-weight hits are allowed when the total \(W\) is positive: those hits simply do not raise the running sum. If \(m=0\) or \(m=N\), the miss or hit normalization is not defined as written, so the gene set must be checked rather than assigned a leading edge from this formula.

Comparing Subsets and Checking the Result

A leading-edge comparison is most useful when the analysis is held constant: the same ranked list, gene identifiers, gene-set version, and weighting convention. For two subsets \(L_A\) and \(L_B\), report their sizes and their intersection, not just a visual impression of overlap. The Jaccard index \(|L_A\cap L_B|/|L_A\cup L_B|\) is one descriptive summary; it does not adjust for pathway size, gene prevalence, or dependence among pathways. In particular, overlapping pathways can naturally have overlapping leading edges, so repeated genes are not independent confirmations of separate biological findings.

Before interpreting a result, inspect the ranked list, pathway membership, extremum, and running-sum trace. Check that the identifiers were mapped consistently, that duplicates were handled by a stated rule, and that the set has both hits and misses in the analyzed list. A useful arithmetic check is that \(R(N)=0\), subject to the stated conditions. Also check whether several ranked statistics are tied or nearly tied around the extremum: small changes in ordering can change which genes fall on one side of the boundary. A zero-weight hit near the boundary may belong to the positional leading edge but make no direct contribution to the score.

A leading-edge subset can be sensitive to the ranking statistic, weighting exponent, gene-set definition, and rank order. It is not a complete list of all genes biologically relevant to a pathway, and genes outside the subset are not thereby shown to be unimportant. Nor does membership establish differential expression, an individual gene’s statistical significance, or causal involvement. Statistical support for the pathway-level result must be assessed using the applicable GSEA inference and multiplicity procedures; the subset itself is not a replacement for them.

Common Mistakes and Alternatives

Wrong: “Every pathway hit raises the weighted running sum.” Correction: A positive-weight hit raises it, a zero-weight hit leaves it unchanged, and a miss lowers it.
Wrong: Treating every member of a significant pathway as a leading-edge gene. Correction: Apply the positional definition relative to the selected extremum.
Wrong: Saying that all leading-edge genes directly caused the extremum. Correction: For a negative ES, misses lower the sum; the negative-direction leading edge identifies pathway members from the trough toward the bottom of the list.
Wrong: Treating overlap across enriched pathways as independent confirmation. Correction: Report shared genes and interpret the overlap in light of pathway overlap and dependence.
Wrong: Reading leading-edge membership as a gene-level significance result or causal conclusion. Correction: It is a descriptive summary of a ranked-list enrichment pattern.

The main alternative is to report the pathway-level ES and its inferential results without emphasizing a subset, or to report all pathway members rather than only those on the relevant side of the extremum. The score is compact but hides which pathway genes occur before or after the extremum; all-member reporting is broader but does not distinguish the genes associated with that particular running-sum boundary. These summaries answer different descriptive questions. None alone identifies a causal mechanism.

Key takeaway: Locate the selected running-sum extremum, then take pathway members on the appropriate side of it: up to the maximum for positive enrichment, or from the minimum onward for negative enrichment. Interpret that subset as a rank- and method-dependent descriptive summary, not as a separate test or causal explanation.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

For \(N=10\) and \(m=2\), what is the decrement at each miss? If a hit has zero weight but the total hit weight is positive, what does that hit do to the running sum?
A negative ES has its minimum at position 6 in a ranked list. A pathway has members at positions 2, 6, and 9. Which members belong to the negative leading edge under the positional definition?
Two leading-edge sets have sizes 4 and 5, with an intersection of 2 genes. What is their Jaccard index?
Why is it incorrect to infer that every gene in a negative leading-edge subset directly pushed the running sum downward?

Solutions. (1) There are \(N-m=8\) misses, so each decrement is \(-1/8=-0.125\). The zero-weight hit leaves the sum unchanged. (2) The members at positions 6 and 9 are in the subset; the member at position 2 is not. (3) The union has \(4+5-2=7\) genes, so the Jaccard index is \(2/7=0.2857\). (4) Hits raise the sum when their weight is positive and leave it unchanged when their weight is zero; misses lower it. The negative leading edge is defined by pathway-member position at or after the trough, not by a claim that each such hit caused the downward movement.