Detecting Batch-Driven Gene-Level Signals

Why Batch-Level Diagnostics Matter

A gene-level rare-variant analysis often reduces many variant calls to an indicator: did a participant carry at least one qualifying variant in the gene? That summary is useful, but it can conceal differences in sequencing, processing, or call quality. If those technical differences vary across batches, they can change which variants are detected—or falsely called—and produce a case-control contrast that resembles a biological association.

A batch is a group of samples processed together under a particular sequencing or laboratory workflow. To assess whether a gene-level result may be batch-driven, examine both the carrier rates and the callability that underlie them. Then compare cases and controls within batches, rather than relying only on a pooled contrast. These checks diagnose possible artifacts; they do not, on their own, prove that a result is technical or establish that it is biological.

Definition: For a specified gene and a fixed set of qualifying-variant rules, a participant is a carrier if at least one qualifying variant is called in that gene. A participant is callable for the gene if the available sequencing data meet the study’s prespecified criteria for making a reliable carrier or noncarrier determination. In batch $b$ and group $g$, the gene-level carrier rate is the number of callable carriers divided by the number of callable participants: $$ \hat{p}_{gb}=\frac{C_{gb}}{N_{gb}}, $$ where $C_{gb}$ is the number of callable carriers and $N_{gb}$ is the number of callable participants in that group and batch.

A noncallable participant is not a noncarrier: the data do not support either conclusion. The exact callability rule depends on the analysis, but it must be defined consistently and be adequate for the gene-level carrier determination. Useful summaries include, for each batch and case-control group, the number enrolled, the number callable, the callable fraction, and the carrier count among those callable. It can also be informative to examine callable regions or variant-level callability, because a single gene-level percentage may hide differences in which parts of the gene could be assessed.

Compare Case and Control Rates Within Each Batch

Start by laying out cases and controls separately within each batch. A pooled carrier-rate difference can reflect a real within-batch difference, a difference in the mix of batches between cases and controls, or both. When the case and control groups have different batch compositions, pooling can create a contrast even if their carrier rates are identical within every batch.

Worked Example: A Pooled Contrast Created by Batch Composition

Suppose a hypothetical study has two batches, with 100 cases and 100 controls overall. Assume everyone is callable. In batch A, there are 20 cases with 1 carrier and 80 controls with 4 carriers. In batch B, there are 80 cases with 8 carriers and 20 controls with 2 carriers. Compare the within-batch and pooled carrier rates.

Batch A: the case rate is $1/20=0.05$, and the control rate is $4/80=0.05$. The within-batch difference is $0.05-0.05=0$.

Batch B: the case rate is $8/80=0.10$, and the control rate is $2/20=0.10$. Again, the within-batch difference is $0.10-0.10=0$.

After pooling, the case rate is $(1+8)/100=0.09$, while the control rate is $(4+2)/100=0.06$. The pooled difference is $0.09-0.06=0.03$, or 3 percentage points. The pooled odds ratio is $$ \frac{9/91}{6/94}=\frac{9\times94}{91\times6}=1.549. $$ The pooled result suggests more carriers among cases, despite equal case and control rates within each batch. Here, cases are more concentrated in batch B, where the carrier rate is higher in both groups. The contrast is therefore compatible with a batch-composition artifact; it does not demonstrate a case-specific excess within either batch.

The arithmetic behind this pattern is a weighted average. Let $r_{gb}$ be the carrier rate in group $g$ and batch $b$, and let $w_{gb}$ be the fraction of callable participants in group $g$ who belong to batch $b$. The pooled rate is $$ \hat{p}_{g}=\sum_b w_{gb}r_{gb}. $$ Even if $r_{\text{case},b}=r_{\text{control},b}$ in every batch, the pooled rates can differ when the batch weights differ. In the example, the case weights are $0.20$ for A and $0.80$ for B, while the control weights are $0.80$ for A and $0.20$ for B. Thus the pooled rates are $0.20(0.05)+0.80(0.10)=0.09$ for cases and $0.80(0.05)+0.20(0.10)=0.06$ for controls. This is why batch-specific contrasts are essential when batch and case status are associated.

Check Callability Before Interpreting Carrier Rates

Carrier rates are interpretable only against an appropriate denominator. If a sample is not callable for the gene, counting it as a noncarrier implicitly assumes that no qualifying variant is present. That assumption is unsupported when the sequencing data cannot establish the sample’s status. Unequal callability between cases and controls can therefore manufacture or obscure a contrast.

Worked Example: Unequal Callability Changes the Apparent Contrast

In a hypothetical batch, 100 cases and 100 controls are enrolled. All 100 cases are callable for a gene, and 10 are carriers. Only 50 controls are callable, and 5 of those are carriers. The remaining 50 controls are noncallable. Compare the callable-only contrast with the result from incorrectly treating noncallable controls as noncarriers.

Among callable participants, the case rate is $10/100=0.10$, and the control rate is $5/50=0.10$. The callable-only odds ratio is $$ \frac{10/90}{5/45}=\frac{10\times45}{90\times5}=1. $$ There is no case-control difference in the observed carrier rates among callable participants.

If the 50 noncallable controls are incorrectly recorded as noncarriers, the apparent control rate becomes $5/100=0.05$. The resulting odds ratio is $$ \frac{10/90}{5/95}=\frac{10\times95}{90\times5}=2.111. $$ That apparent excess among cases is created by using the wrong control denominator. The callability fractions are $100/100=1.00$ for cases and $50/100=0.50$ for controls, an important warning in its own right. The callable-only comparison does not resolve why half the controls are noncallable; it shows why their status must not be silently coded as noncarrier.

A callability summary should be read alongside carrier counts, not as a substitute for them. Similar overall callable fractions do not guarantee that the same bases or variants were assessable in cases and controls. Conversely, a callability difference is a warning requiring investigation, not automatic proof that the association is an artifact. Record the criteria used to define callability, apply them consistently, and make the callable denominators visible in the analysis.

Use Stratified Contrasts to Assess Persistence

When each batch contains both cases and controls, calculate the case-control contrast separately in each batch. A consistent direction and broadly similar magnitude across batches is more compatible with a result that is not solely due to different batch composition. Markedly different contrasts, or a pooled association that disappears within batches, signal that batch structure may be driving the result. Sparse rare-variant counts make these comparisons noisy, so a zero or unstable estimate in a small batch should not be treated as definitive evidence of no effect.

Worked Example: A Case-Control Contrast Within Batches

Consider a hypothetical gene with complete callability in two batches. In batch A, 10 of 50 cases and 5 of 50 controls are carriers. In batch B, 2 of 50 cases and 1 of 50 controls are carriers. Calculate the rate differences and a common batch-stratified odds ratio.

In batch A, the case and control rates are $10/50=0.20$ and $5/50=0.10$, giving a difference of $0.10$, or 10 percentage points. In batch B, the rates are $2/50=0.04$ and $1/50=0.02$, giving a difference of $0.02$, or 2 percentage points. Both observed differences are positive, although the batch-specific counts are small.

For a common stratified odds ratio, let $a_b$ and $b_b$ be the carrier and noncarrier counts among cases, and $c_b$ and $d_b$ the corresponding counts among controls. With $n_b=a_b+b_b+c_b+d_b$, the Mantel-Haenszel estimate is $$ \widehat{\mathrm{OR}}_{\mathrm{MH}}= \frac{\sum_b a_bd_b/n_b}{\sum_b b_bc_b/n_b}. $$ For batch A, $a_A=10$, $b_A=40$, $c_A=5$, $d_A=45$, and $n_A=100$. For batch B, $a_B=2$, $b_B=48$, $c_B=1$, $d_B=49$, and $n_B=100$. Substitution gives $$ \widehat{\mathrm{OR}}_{\mathrm{MH}}= \frac{(10\times45)/100+(2\times49)/100} {(40\times5)/100+(48\times1)/100} =\frac{4.50+0.98}{2.00+0.48} =\frac{5.48}{2.48}=2.210. $$ The stratified estimate indicates higher carrier odds in cases after accounting for these two batch strata. It is not proof of a biological effect: technical differences may still operate within batches, and the small counts imply substantial sampling uncertainty. The estimate also summarizes a common association and should be interpreted cautiously if the batch-specific effects are genuinely very different.

A Practical Diagnostic Sequence

Fix the gene-level definition.
Specify the qualifying variants and the carrier rule before comparing groups. Use the same rules for cases and controls.

Tabulate callability.
For each batch and case-control group, report enrolled and callable counts, callable fractions, and the basis for calling a participant callable for the gene.

Calculate batch-specific carrier rates.
Use callable participants as the denominator. Show the numerator and denominator, since rates based on very few carriers can be unstable.

Compare pooled and within-batch results.
Check whether a pooled case-control difference is also present within batches, and whether its direction is reasonably consistent across batches.

Interpret the pattern as a diagnostic.
Investigate conspicuous differences in callability or rates as potential technical concerns; do not treat the pattern alone as proof of artifact or causation.

Common Pitfalls and Interpretation

A batch-driven signal can arise because batches differ in technical properties that affect variant detection, while cases and controls are distributed unevenly across those batches. But a batch-specific rate difference is not automatically technical: chance variation, population composition, or genuine differences in the sampled participants can also contribute. Likewise, stable within-batch contrasts do not guarantee that every technical bias has been removed. The diagnostic question is whether the evidence is compatible with a technical explanation and whether the association remains visible under comparisons that respect the batch structure.

Wrong: “The pooled case carrier rate is higher, so the gene is associated with case status.” Correction: First check batch composition and compare cases with controls within batches; differing batch weights alone can create a pooled contrast.
Wrong: “A noncallable sample has no qualifying variant.” Correction: Noncallable means status is unresolved under the stated criteria. Do not count it as a noncarrier; report callability and use an appropriate analysis.
Wrong: “A batch difference proves a sequencing artifact.” Correction: It identifies a possible source of bias to investigate. Batch summaries diagnose patterns, but do not determine their cause.
Wrong: “Similar overall callability means equal variant detection.” Correction: The callable regions and variant-level performance may still differ. Review the callability definition and relevant summaries for the gene.

Key takeaway: Interpret a gene-level case-control signal alongside batch-specific carrier numerators and callable denominators. A pooled association that vanishes within batches, or coincides with uneven callability, deserves technical scrutiny; neither pattern alone establishes the cause.

Check Your Understanding

Use the denominators and batch structure explicitly in your answers.

Why can pooled case and control carrier rates differ even when their rates are equal within every batch?
A control group has fewer callable participants than the case group. Why is it incorrect to code all noncallable controls as noncarriers?
What quantities should be reported to make a batch-specific carrier rate interpretable?
If a pooled association disappears after comparing cases and controls within each batch, what concern does that raise, and what does it not prove?
Why can similar gene-level callable fractions fail to establish that cases and controls had comparable opportunity to detect every qualifying variant?