Batch Effects in Genotyping QC

Why Compare Genotyping Metrics by Batch?

Genotyping data can differ across plates, laboratories, centers, or sequencing runs even when the underlying biological samples are comparable. A plate may have unusually high missingness, one sequencing run may produce lower genotype-quality scores, or a locus may show a striking allele-frequency difference between centers. These patterns matter because technical differences can affect downstream results. But a difference between batches is a signal to investigate, not by itself proof of a technical error.

A useful batch comparison asks three related questions: are genotypes being called at similar rates, do allele-frequency summaries look unexpectedly different, and are genotype-quality scores comparable? The comparison is most informative when batches were processed with the same assay and variant definitions, the same metrics are calculated in the same way, and differences in the samples being processed are considered.

Choose Comparable Metrics and Denominators

Call rate is the fraction of genotype opportunities that produced a call. For a set of samples at a variant, it is a variant call rate; for a sample across a set of variants, it is a sample call rate. Report which quantity is being compared, and make the denominator explicit. A comparison is not meaningful if one batch is evaluated over a different set of variants, or if failed or excluded calls are handled differently.

For example, if 8,900 of 10,000 eligible sample-by-variant entries are called, the call rate is \(8{,}900/10{,}000=0.890\), or 89.0%. When comparing batches, calculate sample-level call rates and variant-level call rates as well as an overall summary. An overall percentage can conceal a small number of very poor samples or loci.

Allele-frequency comparisons should be made at the same normalized variant and with the same allele coding. For a simple diploid example with no missing calls, the frequency of a specified allele is its number of observed copies divided by the number of allele copies. With missing calls, the denominator is the number of allele copies actually observed. The frequency is a descriptive summary, not a diagnostic verdict: genuine differences in ancestry or sample composition can change it.

Genotype quality (GQ) summarizes confidence in a genotype call. In common formats it is Phred-scaled, so larger values indicate greater confidence and smaller values indicate less confidence; the exact definition and calibration depend on the calling software. Compare GQ distributions among called genotypes, and show call rates alongside them. Looking only at GQ among the calls that survived can hide a batch where many difficult genotypes were never called.

Worked Examples

Plots That Reveal Batch-Associated Patterns

Plots help determine whether a difference is widespread, localized, or driven by a few observations. Select a plot that matches the unit of analysis and retain batch labels throughout.

• Sample call-rate boxplots or dot plots by plate or run: a broadly lower distribution in one batch suggests a group-level shift; a handful of isolated low points suggests that the problem may be concentrated in individual samples.
• Variant call-rate plots by batch: compare per-variant call rates across batches, or plot one batch against another. A broad shift across many variants differs from a small set of loci with poor performance.
• Allele-frequency scatterplots: plot the frequency of the same allele in one batch against its frequency in another. Points far from the equality line identify loci for review. A shift across many loci can reflect sample composition, coding, or processing differences; it does not identify which explanation is correct.
• GQ boxplots, histograms, or density plots: compare both the center and spread of GQ among called genotypes. A lower distribution may indicate reduced call confidence, but should be interpreted alongside call rates and software settings.
• Plate-position plots or heatmaps: show a sample-level metric at each physical plate position. A spatial pattern, such as poorer call rates along one edge, can point to a location-related process issue that a single plate-wide average would obscure.

Use consistent axes and definitions across groups. For instance, independently scaled axes can make substantially different distributions look alike. Label the number of samples or calls behind each summary, since a group with few observations may produce unstable-looking summaries. A plot should support a focused follow-up, not replace checking the underlying records and processing metadata.

A Practical Comparison and Its Limits

A disciplined comparison proceeds from data definitions to summaries and then to investigation. First confirm that batch labels are accurate and that each batch was evaluated on comparable variants, with consistent genotype coding and missing-call rules. Then calculate sample and variant call rates, summarize allele frequencies at shared variants, and compare GQ distributions. Plot the measurements at the level where the problem could arise: sample, variant, or plate position.

One small calculation explains why an overall call rate can be misleading. If batch \(b\) has \(N_b\) eligible genotype opportunities and call rate \(r_b\), then its number of calls is \(N_b r_b\). Across batches, the pooled rate is therefore

This is a denominator-weighted average, not generally the ordinary average of the batch rates. For example, suppose Batch A has 100 opportunities and a call rate of 0.90, while Batch B has 300 opportunities and a call rate of 0.98. The pooled rate is \((100(0.90)+300(0.98))/(100+300)=(90+294)/400=0.960\). The unweighted average is \((0.90+0.98)/2=0.940\), which gives the small batch the same weight as the batch with three times as many opportunities. Reporting the pooled value alone would also hide the 8-percentage-point difference between the batch rates, so report both group summaries and their denominators.

Common Mistake: Treating Every Difference as a Technical Artifact

A common error is to see different allele frequencies across centers and conclude that one center genotyped the locus incorrectly. That conclusion does not follow from the frequency comparison alone. The groups may differ in ancestry or other biological composition, and an allele-frequency contrast can be genuine. Conversely, similar overall frequencies do not rule out a technical problem affecting call rate or genotype quality.

The correction is to treat batch-stratified plots as diagnostic evidence, check whether the comparison uses the same samples, variants, coding, and metric definitions, and consider plausible technical and biological explanations. When groups have different sample sizes, report their denominators; when calls are missing, make clear which observations contribute to a frequency or GQ summary. A batch-associated difference can justify follow-up without establishing either causation or a particular error mechanism.