Integrating Evidence for Candidate-Gene Ranking

Why Integrate Evidence for Candidate Genes?

A locus can contain several plausible genes, and different evidence types can point in different directions. A coding consequence may favor one gene, enhancer-to-gene links another, and molecular association evidence a third. A useful ranking makes these disagreements visible rather than hiding them inside an unexplained judgment.

A ranking is a structured way to order candidates using the evidence available for a locus. It is not, by itself, a test of a gene-level hypothesis or a probability that a gene is causal. Its result depends on the candidate list, how evidence is summarized, and the rules used to combine it. The goal is therefore not to produce a universally correct numerical score, but to make the reasoning explicit enough to inspect, compare, and revise.

Separate the Evidence Before Combining It

Begin by defining the candidate set: the genes eligible to be ranked at the locus. The set should follow a stated inclusion rule, such as genes in a specified genomic interval or genes supported by a prespecified mapping procedure. Changing the candidate set can change the top-ranked gene, even when the evidence for each gene is unchanged.

For each candidate, summarize evidence in distinct categories. For example, a team might use a coding component \(C_g\), a regulatory-link component \(R_g\), and an association component \(A_g\) for gene \(g\). Each component should have a documented scale and interpretation. A value of 1 might represent the strongest support under a rubric, while 0 represents no support under that rubric. These values are evidence summaries, not automatically calibrated probabilities.

• Coding evidence can summarize the consequence and association support for variants that affect the gene’s coding sequence.
• Regulatory evidence can summarize functional evidence and the strength of links from regulatory elements to the gene.
• Association evidence can summarize gene-specific molecular association evidence that is relevant to the locus, such as a molecular QTL result concordant with the GWAS signal.

A locus-wide GWAS association establishes that the region merits attention; if that same association statistic is identical for every candidate gene, it cannot distinguish the genes. Gene ranking requires evidence that differentiates candidates, such as a coding consequence in one gene, a credible-variant-linked regulatory element, or gene-specific molecular association evidence. A small \(p\)-value for the locus should not simply be copied into every gene’s score or treated as proof for the nearest gene.

Before combining components, specify how they were constructed. If two components summarize the same assay or the same underlying evidence, adding both at full weight can count that evidence more than once. The earlier material on correlated functional annotations addresses this issue; for a ranking, the practical requirement is to disclose shared sources and either combine redundant measures first or use weights that reflect the overlap.

A Transparent Weighted Score

One simple scoring rule is a weighted sum. Let \(w_C\), \(w_R\), and \(w_A\) be nonnegative weights that sum to 1. The score for gene \(g\) is:

The weights express the relative influence assigned to each evidence category; they are not estimated effect sizes unless a separate training and calibration procedure has been carried out. If all components are on a comparable 0-to-1 scale, \(S_g\) will also be on that scale. A score of 0.7 means that the weighted evidence summary is 0.7 under the chosen definitions and weights. It does not mean there is a 70% probability that the gene is causal.

This rule is most interpretable when the components use consistent scoring rubrics, the candidate set is defined in advance, and weights are chosen before examining which gene wins. If evidence is missing, do not silently substitute 0 unless 0 has explicitly been defined to mean “measured and no support.” Missing evidence means the component was not observed or evaluated; that is different from a measured negative result.

Rule-Based Prioritization

A rule-based ranking can be preferable when evidence categories are not naturally measured on comparable numerical scales, when a decisive evidence class has priority, or when stakeholders need a clear decision rule. A rule must still be specified precisely enough that another analyst could apply it to the same evidence.

For example, a prespecified hierarchy might rank genes in this order: first, genes with a high-confidence coding consequence on a strongly association-supported variant; second, genes with a strong regulatory link to an association-supported variant; third, genes with gene-specific molecular association evidence; and finally, genes supported only by genomic proximity. Ties within a tier could be broken by a documented secondary criterion. This is not the same as informally choosing whichever gene “looks most plausible”: the evidence categories, thresholds, and tie-breaker are explicit.

A hierarchy makes a strong priority clear but can be insensitive to evidence outside the highest applicable tier. A weighted sum allows several moderate evidence types to accumulate, but can make rankings depend substantially on the chosen weights. Neither approach is inherently superior. The appropriate choice depends on the scientific question and how defensible the required assumptions are.

Why Weight Choices Matter

A weighted ranking can change when weights change. A short derivation shows exactly how two genes compare. For genes \(g\) and \(h\), subtract their scores:

Thus, \(g\) outranks \(h\) precisely when the right-hand side is positive. If \(g\) has a coding advantage but \(h\) has regulatory and association advantages, the sign depends on how heavily each component is weighted. This identity follows directly by subtracting the two weighted-sum formulas and grouping terms by component; it also provides a simple way to identify which weights could reverse an ordering.

Assumptions, Reporting, and Common Pitfalls

A score should be treated as a disciplined summary, not as a substitute for examining the underlying evidence. At minimum, report the candidate set, component definitions, scoring scales, weights or decision rules, and the evidence that determined the leading rank. If a score is used for decisions beyond descriptive prioritization, its performance should be evaluated on appropriate data without leakage from the evaluation set into score construction.

The evidence components also need to be interpreted in context. A strong regulatory link is not automatically evidence that the linked gene mediates the GWAS association. A molecular association can support a gene without proving the causal direction or mechanism. A coding consequence can be compelling, but its relevance still depends on whether the variant is credibly associated with the locus and whether the consequence is biologically interpretable. Combining evidence does not remove these limitations.

• Common mistake: using the locus \(p\)-value as a gene score. The reasoning is that a very small locus \(p\)-value makes every nearby gene highly supported. The correction is that a shared regional association does not distinguish those genes; gene-specific evidence must connect the signal to a candidate.
• Common mistake: interpreting a score as a causal probability. The reasoning is that a score between 0 and 1 looks like a probability. The correction is that a weighted score has the meaning assigned by its component scales and weights; it is not a probability unless separately calibrated and validated as one.
• Common mistake: counting the same evidence twice. The reasoning is that each annotation track provides another independent reason to raise a score. The correction is to check whether tracks share assays, variants, or underlying measurements before treating their contributions as distinct.
• Common mistake: treating unavailable evidence as negative evidence. The reasoning is that a blank component can be entered as zero for convenience. The correction is to distinguish an unmeasured component from an observed result indicating no support, and state how missing information is handled.