Reading the Report Before Reading the Flags
A genotype QC summary compresses many checks into a small set of counts, rates, and flags. Its purpose is not to declare every sample or variant “good” or “bad.” It is to show where the data need review and to make the evidence behind later filtering decisions visible. A low call rate, a small Hardy-Weinberg p-value, or a relatedness flag each describes a particular feature; none alone gives a complete account of data quality.
Begin by identifying the unit of each row or count: sample, variant, or pair of samples. Then check the analysis set and denominators. For example, a variant call rate among all enrolled samples is not directly comparable to one calculated among samples remaining after exclusions. Likewise, a Hardy-Weinberg test among unrelated controls answers a different QC question from a test that includes cases or close relatives.
A report is interpretable only if it defines what counts as eligible and called. Check whether it includes autosomal variants only, whether it excludes low-frequency or otherwise ineligible variants, how sex chromosomes and ploidy are handled, and which samples are included in each calculation. A rate can be numerically precise yet misleading if its denominator is unclear.
Call Rates: Verify the Denominators
For sample \(i\), let \(M_i\) be the number of eligible genotype entries and \(C_i\) the number called. The sample call rate is \(C_i/M_i\). For variant \(j\), if \(N_j\) samples are eligible and \(G_j\) have a called genotype, the variant call rate is \(G_j/N_j\). These rates summarize missingness along different directions of the genotype matrix: one scans across variants for a sample, the other across samples for a variant.
A mean of sample call rates and an overall call rate need not be the same if samples have different numbers of eligible entries. The overall rate weights each sample by its denominator. The same principle applies to aggregating variant rates. The following calculation shows why the distinction matters.
Worked Example: Why an Overall Call Rate Is Weighted
Suppose a hypothetical dataset has two samples. Sample A has 98,000 called genotypes among 100,000 eligible entries, and sample B has 891,000 called genotypes among 900,000 eligible entries. Calculate the two sample rates, their unweighted mean, and the overall rate.
Sample A’s rate is \(98{,}000/100{,}000=0.9800\), or 98.00%. Sample B’s rate is \(891{,}000/900{,}000=0.9900\), or 99.00%. Their unweighted mean is \((98.00\%+99.00\%)/2=98.50\%\).
Across both samples there are \(98{,}000+891{,}000=989{,}000\) called entries and \(100{,}000+900{,}000=1{,}000{,}000\) eligible entries. Thus, the overall rate is \(989{,}000/1{,}000{,}000=0.9890\), or 98.90%. It is not the unweighted mean because Sample B contributes nine times as many eligible entries as Sample A.
A discrepancy between a reported overall rate and the mean of sample rates is not automatically a software error. First determine which summary was reported and whether the denominators differ.
Hardy-Weinberg, Heterozygosity, Sex, and Relatedness
A Hardy-Weinberg (HWE) result is usually reported as a count of variants below a p-value threshold, sometimes with the smallest p-values listed. Interpret it with the test population, variant set, and test method in view. For QC, the tests are commonly restricted to an appropriate set of samples, such as unrelated controls, and to variants for which the test is meaningful. A small p-value is evidence of departure from the model used by the test; it is not the probability that the variant is erroneous, nor does it identify the cause of departure. Genotyping error, population structure, relatedness, selection, or other features can contribute. The earlier material on interpreting Hardy-Weinberg failures gives the underlying test interpretation; a summary report requires checking whether its analysis set supports that interpretation.
Heterozygosity summaries describe how often heterozygous genotypes occur across a specified set of variants. An outlier flag generally means that a sample’s value is unusually high or low relative to a reference distribution, often using a stated standardized measure. Verify which variants formed that reference set and whether ancestry, variant filtering, or batch differences could shift the distribution. An outlier is a diagnostic signal, not a direct measurement of contamination, inbreeding, or any single cause.
A sex-check flag reports disagreement between recorded information and a genomic inference based on sex-chromosome data. Interpret the exact inference and reference categories used, and remember that unusual ploidy, chromosome variation, sample mix-ups, or limitations in the recorded information may require investigation. The flag alone does not establish a person’s sex or gender. A relatedness flag identifies a pair whose estimated genetic similarity exceeds a specified criterion. It is a pair-level result: it does not, on its own, say which sample to retain or whether family data should be excluded.
| Report item | What it summarizes | What to check before acting |
|---|---|---|
| Sample call rate | Missingness across eligible variants for a sample | Eligible variant set, denominator, batch pattern, and sample-level failures |
| Variant call rate | Missingness across eligible samples for a variant | Eligible samples, localized missingness, and whether the variant is technically problematic |
| HWE results | Evidence of departure from the specified HWE model at tested variants | Test population, ancestry and relatedness, variant eligibility, and multiplicity context |
| Heterozygosity outliers | Samples unusually high or low on the reported heterozygosity measure | Variant set, reference distribution, and other sample or batch evidence |
| Sex discordance | Disagreement between recorded information and the stated genomic check | Inference method, chromosome data quality, and appropriate confidential review |
| Relatedness flags | Sample pairs exceeding a stated relatedness criterion | Estimator, threshold, intended study design, and consequences of excluding either sample |
Worked Examples from a Compact Summary
Worked Example: Interpreting a Hypothetical Project Summary
A hypothetical report covers 500 samples and 700,000 eligible autosomal variants. It lists an overall genotype call rate of 99.10%, a median sample call rate of 99.42%, a minimum sample rate of 96.80%, a median variant call rate of 99.70%, and a minimum variant rate of 91.20%. It also reports 3,210 variants with HWE p-values below \(10^{-6}\) in unrelated controls, eight heterozygosity outliers, four sex-check discordances, and 12 pairs above a stated kinship threshold.
The overall 99.10% rate indicates that 99.10% of eligible genotype entries were called, under the report’s definition. The minimum sample rate warrants inspection of the affected sample, but the median of 99.42% shows that the minimum does not describe a typical sample. Similarly, the minimum variant rate indicates at least one variant with substantial missingness, but the median variant rate is much higher. The summary does not reveal whether missingness is concentrated by batch, phenotype, or another grouping; those patterns should be checked rather than inferred from the overall rates.
The 3,210 HWE results are a count below a specified threshold, not 3,210 confirmed genotyping errors. Interpretation depends on the stated control set and test assumptions. The eight heterozygosity outliers, four discordances, and 12 related pairs are different kinds of flags with different units. They should not be added into a single “number of bad samples”: a sample can appear in more than one category, and relatedness counts pairs rather than individuals.
A sound first reading is therefore: most entries are called, with some low-rate exceptions; several HWE tests warrant variant-level review; and a limited number of sample or pair diagnostics need follow-up. The summary alone does not justify removing every flagged item.
Worked Example: Reading Overlapping Sample Flags
In a hypothetical dataset, Sample S17 has a sample call rate of 96.80%, a high heterozygosity flag, and a sex-check discordance. Sample S42 has a call rate of 99.60% and is the other member of a relatedness-flagged pair with S17. What can be concluded from these entries?
S17 has three distinct warning signals: elevated missingness relative to the project’s stated criterion, an unusual heterozygosity value, and disagreement in the specified sex check. Their overlap makes S17 a high-priority sample for review, but it does not establish one cause. For example, a technical problem could affect several metrics, but the report does not prove that explanation. The investigator should inspect the definitions, relevant plots or diagnostics, and any available sample-level provenance before deciding what to do.
The relatedness flag concerns the S17–S42 pair, not simply S17. S42’s call rate is high, but that fact does not by itself decide which member of the pair, if either, is appropriate to retain. The study design matters: the pair may be expected family data, or one sample may be a duplicate. The report provides evidence to investigate these possibilities; it does not settle them.
Worked Example: Interpreting a Variant Flag Alongside Its Call Rate
A hypothetical variant has a 99.80% call rate and an HWE p-value of \(2.0\times10^{-8}\) in the report’s unrelated control set. The report’s stated HWE review threshold is \(10^{-6}\). Interpret both values.
The call rate is \(0.9980\), so 99.80% of eligible samples have a called genotype at this variant. The HWE result is below the stated threshold because \(2.0\times10^{-8}<10^{-6}\), so the variant meets the report’s criterion for HWE review. The small p-value does not contradict the high call rate: a genotype can be called for nearly everyone and still show a genotype pattern that departs from the test model.
The next step is to confirm that the controls and variant satisfy the report’s intended test conditions, then examine relevant variant-level evidence, such as genotype quality and whether results vary by batch or ancestry group. The p-value is not \(P(\text{variant is erroneous}\mid\text{data})\), and the call rate is not evidence that the HWE departure is harmless.
Assumptions, Diagnostics, and Common Pitfalls
A summary should state enough information to assess the conditions behind its results. Call rates require well-defined eligibility and consistent missing-call coding. HWE tests require a defined test population and a test appropriate to the variant and analysis set; departures may have multiple explanations. Heterozygosity outlier comparisons require a clearly described reference set. Sex checks require a stated genomic inference and usable chromosome data. Relatedness flags require a named or otherwise documented estimator and threshold. These checks do not guarantee that assumptions hold, but they make the reported result interpretable.
Useful diagnostics look for patterns the compact summary hides. Compare sample call rates by batch and other relevant groups; inspect whether low variant rates cluster in particular samples; and compare HWE flags across analysis subsets when the design supports that comparison. For sample outliers, examine the distribution and the variants used to calculate it. For sex-check or relatedness flags, confirm that identifiers and inputs are correct and consider relevant study-specific explanations. A discrepancy that vanishes under a defensible change of analysis set may be informative, but it should not be resolved by choosing whichever calculation gives fewer flags.
A common mistake is to interpret each threshold as an automatic deletion rule: “It crossed the cutoff, therefore remove it.” The correction is to distinguish screening from decision. A threshold identifies results for review under a chosen rule; the final action also depends on the metric’s assumptions, corroborating evidence, and the study objective. The opposite mistake is to ignore all flags because no single one proves a problem. A flag is not proof, but it is a reason to examine the relevant evidence.
Compared with inspecting raw genotype data alone, a compact report makes broad patterns and exceptional cases easier to find, but it hides distributional detail and dependencies among flags. Compared with a single combined “QC score,” separate metrics preserve the reason for each concern and avoid treating unlike evidence as interchangeable. The report is best used as an index into focused diagnostics, not as a substitute for them. Thresholds may also be project-specific; a threshold chosen for one platform, sample composition, or study design should not be treated as a universal biological boundary.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A sample has 19,600 called genotypes among 20,000 eligible entries. What is its sample call rate, and what does that rate describe?
- Two groups have sample call rates of 98% and 99%, but the second group has five times as many eligible entries. Is the overall rate 98.5%? Explain.
- A variant has an HWE p-value below the report threshold in unrelated controls. Does this establish that the variant is erroneous? Name two relevant checks.
- A sample is both a heterozygosity outlier and a member of a relatedness-flagged pair. Why should those flags not be collapsed into one conclusion?
Solutions. (1) \(19{,}600/20{,}000=0.9800\), or 98.00%. It describes the fraction of that sample’s eligible genotype entries that are called. (2) No. The overall rate is weighted by eligible-entry counts: with 20 eligible entries in the first group and 100 in the second, the combined rate is \((0.98\times20+0.99\times100)/120=0.9883\), or about 98.83%. (3) No. The p-value indicates evidence of departure from the specified test model, not the probability of error. Check that the test population and assumptions are appropriate, and examine variant-level quality or patterns across batches or ancestry groups. (4) The outlier is a sample-level result about heterozygosity; the relatedness flag concerns a pair. Neither identifies a single cause or dictates which sample to retain, so each requires its own contextual review.