Reporting Fine-Mapping Results Transparently

Why transparent reporting matters

A fine-mapping result is conditional on a defined region, a set of candidate variants, a statistical model, prior assumptions, and linkage disequilibrium (LD) information. A list of variants with large posterior inclusion probabilities (PIPs) is therefore not a complete result. Readers need enough information to understand what those probabilities refer to, how a credible set was formed, and whether reasonable changes in assumptions alter the conclusion.

Transparent reporting does not make a model assumption-free. It makes the assumptions visible and the analysis reproducible. In particular, a posterior probability is a probability under the stated model and priors; it is not a direct measurement of biological causality. The earlier material on prior sensitivity analysis and LD sensitivity analysis provides the relevant framework for reporting how these choices affect results.

Define the analysis before presenting its results

The region must be identifiable without ambiguity. Give the chromosome, start and end coordinates, coordinate convention, and genome assembly, such as GRCh38. State how the region was selected and whether boundaries were expanded or otherwise changed during analysis. Identify the phenotype, the association-study sample or summary-statistics source, and relevant sample characteristics, including ancestry. If multiple ancestries, cohorts, or phenotypes were combined, describe that combination rather than reporting a single undifferentiated result.

Describe the candidate-variant set, not just the region boundary. Report the number of variants considered, the variant identifiers and alleles, and any exclusions or harmonization rules that could change which variants are represented. A variant absent from the analyzed set cannot receive a reported PIP. Where variant identifiers or allele alignment matter, provide a machine-readable table rather than relying only on prose.

State the model at the level needed to interpret its outputs: whether it allows zero, one, or multiple causal variants; how signals or causal configurations are represented; and the assumed relationship between association evidence and LD. Also report the prior specification, including the prior probability or expected number of causal variants, any variant-specific prior weights, and any effect-size prior or other model-specific parameters. If software supplies default values, identify the software and version and give the values actually used. “Default settings” alone is not a sufficient description.

These details matter because the same data can yield different posterior summaries under different models or priors. A report should not imply that one specification is the only possible model. It should make clear which specification generated the results being shown and direct attention to sensitivity results when these are available.

Report probabilities and credible sets precisely

In a model that permits multiple causal variants, marginal PIPs need not sum to one: their sum is the posterior expected number of included variants. By contrast, probabilities used to rank candidates for a particular signal are often normalized across variants conditional on that signal being present. The report must say which quantity it presents. This distinction is especially important when a method also assigns probability to a null or no-signal model. A set constructed from probabilities conditional on a signal should not be described as accounting for uncertainty that the signal exists unless that uncertainty is included in the calculation.

A credible set for a signal is a collection of candidate variants whose cumulative signal-specific posterior probability reaches a stated target, such as 0.95. One common construction ranks variants from largest to smallest signal-specific probability and takes the shortest leading sequence whose cumulative probability reaches the target. Report the target and the construction rule. If the software uses a different rule, describe that rule instead of assuming the common one.

Purity summarizes how strongly the variants in a set are correlated in the LD matrix. A common convention is the minimum pairwise \(r^2\) among distinct variants in the set. State the convention explicitly: some software uses a different summary, such as an absolute correlation or a quantile of pairwise values. A numerical purity value without its definition cannot be compared reliably across analyses. Purity is a property of the set under the specified LD data; it is not a probability that the set contains the causal variant.

How the ranked construction works

Suppose the signal-specific probabilities are \(q_1,\ldots,q_m\), with \(q_j \geq 0\) and \(\sum_{j=1}^{m}q_j=1\). For a target \(c\), sort them in descending order, so \(q_{(1)}\geq q_{(2)}\geq\cdots\geq q_{(m)}\), and choose the smallest \(K\) such that the cumulative sum reaches or exceeds \(c\):

This is a smallest-cardinality set among sets formed from those candidate variants: for any set of \(k\) variants, its total probability cannot exceed the sum of the \(k\) largest probabilities. Consequently, if the leading \(K-1\) probabilities sum to less than \(c\), no set of fewer than \(K\) variants can reach the target. This justification depends on the probabilities being the appropriate normalized probabilities for the signal; it does not make an incorrectly defined or incompletely specified probability suitable for set construction.

Make the LD and sensitivity analysis auditable

LD is not a generic property supplied by the word “reference.” Report the source population or reference panel, sample size if known, ancestry, variant set, and the procedure used to estimate or obtain the LD matrix. Include the genome build and allele harmonization details needed to align LD variants with association data. If LD was estimated in the association sample, say so; if a reference panel was used, identify it and its version or release. Explain how missing or mismatched variants were handled. These items let readers judge whether the LD input plausibly represents the analyzed samples.

Sensitivity results should be concise but concrete. For each alternative prior or LD specification, identify what changed and report the corresponding quantities used to assess the conclusion: for example, PIP changes for leading variants, set membership, set size, or purity. State which choices were held fixed. If a finding changes, describe the change rather than compressing it into “sensitive.” If it does not change, avoid calling that proof of robustness to all possible model errors; the tested alternatives only cover the choices examined.

A compact table can make the main result and alternatives easier to compare. Give each row a clearly identified signal and analysis specification. A reader should be able to distinguish a change in the model from a change in the LD source, and to tell whether PIPs are marginal or signal-specific.

Diagnostics, limitations, and common reporting failures

Before release, check that variant identifiers and alleles align across the association data, LD matrix, and PIP output; that the stated genome build is consistent; and that the reported set membership reproduces from the probabilities and threshold. Check that purity is computed on the members of the set, with the stated definition, and that the LD matrix contains the required variant pairs. When these checks fail, a report may contain plausible-looking numbers that describe different variants or different inputs.

A broad set with low minimum pairwise \(r^2\) can be difficult to summarize as one tightly correlated group. Report that result, rather than hiding it behind a threshold label. Conversely, high purity does not establish that the set contains the causal variant: the conclusion remains conditional on the candidate set, model, priors, and LD input. An omitted causal variant or a poorly matched LD reference can undermine interpretation even when the reported calculations are internally consistent.

Common mistakes include reporting PIPs without defining whether they are marginal or signal-specific; calling a PIP a probability that a variant is biologically causal without qualification; reporting a “95% set” without the construction rule; and giving a purity number without defining the LD summary. Another frequent error is reporting only the primary specification while describing results as robust, with no alternatives or comparison quantities. The correction is to state the conditioning and definitions, provide the needed inputs and settings, and report sensitivity evidence in terms that can be checked.

The main alternative to a detailed narrative is a structured table with accompanying machine-readable files. Tables are efficient for comparing signals and specifications, but only if the columns define the quantities and units clearly. A prose-only report can explain modeling choices, while a table or supplementary file is better for long variant lists. Neither format removes the need to state model, prior, threshold, purity, and LD conventions. Even a fully reproducible report remains limited by the assumptions and data available; transparency supports evaluation, not certainty.