Twins Reared Apart: Measuring and Preventing Overcorrection in Single-Cell Batch Integration

Single-cell RNA sequencing measures gene activity in thousands of individual cells. The data come in "batches," meaning separate runs, labs, or reagent lots, and each batch leaves its own technical fingerprint. Correction methods remove these fingerprints, but nobody can easily tell when they also remove real biology. The biology most at risk is the subtle variation between people, or between cell states, inside a single cell type. Behavioral genetics solved a structurally identical problem decades ago. Studying identical twins raised in different homes lets you separate what the genome does from what the home does, because the genome is held fixed while the environment changes. We propose the same move for single-cell data. When the same donor's cells are sequenced in two different batches, they are "twins reared apart." The donor's biology is fixed while the technical environment changes, so any biology that correction removes shows up as a drop in how well the donor resembles itself across batches.

The proposal has two parts. The first is a cell-level variance-components model, the single-cell analogue of the twin-study decomposition into genetic, shared-environment, and unique-environment parts. Within each cell type it splits expression variation into a donor component, a batch component, and a cell-to-cell component. It fits this on raw counts, so the many zeros caused by molecules that were present but never captured are not counted as biological variation. The second part is a correction method that removes only the estimated batch component, allows that component to differ across cell states, and leaves the donor and cell-state components untouched. The diagnostic is the main contribution. The corrector is a test of whether the diagnostic can guide a better method.

Core Claim

In a design where donors are crossed with batches, meaning at least some donors appear in more than one batch, donor variance and batch variance are separately identifiable within each cell type. That makes "how much real donor and cell-state variance did integration destroy?" a measurable quantity rather than a judgment call. We claim three things. Current integration methods lose a large and method-dependent fraction of within-cell-type donor variance, and neighborhood-mixing metrics do not detect this loss. A count-based decomposition estimates that loss more accurately than one built on log-transformed data, especially in sparse cell types. Removing only the estimated batch component retains more donor variance than learned integrators at comparable batch mixing.

The model, for gene $g$ in cell $c$ of cell type $k$, from donor $d$ and batch $b$, is

$$y_{gc} \sim \mathrm{NB}\!\left(s_c\, e^{\eta_{gc}},\ \phi_g\right),\qquad \eta_{gc} = \mu_{gk} + u_{gd} + v_{gb,z(c)} + w_{gc}.$$

Here $y_{gc}$ is the observed count. $s_c$ is the cell's sequencing depth. $\phi_g$ is the negative-binomial dispersion, which absorbs counting and capture noise. $\mu_{gk}$ is the cell-type mean. $u_{gd}\sim N(0,\sigma^2_{D,gk})$ is the donor effect. $v_{gb,z(c)}\sim N(0,\sigma^2_{B,gk})$ is a batch effect that may depend on a coarse cell-state bin $z(c)$. $w_{gc}\sim N(0,\sigma^2_{W,gk})$ is residual biological cell-to-cell variation.

The overcorrection index for a method is the median over genes of $\hat\sigma^2_{D}$ after correction divided by $\hat\sigma^2_{D}$ before. The before-correction donor variance is identifiable only because donors are crossed with batches. This mirrors the reared-apart twin logic in the pedigree mixed models summarized in the file, where the variance of a trait is split as $\sigma^2_A\mathbf{A}+\sigma^2_C\mathbf{C}+\sigma^2_E\mathbf{I}$.

How This Relates to Prior Work

Four groups of prior work bear on the proposal, each in a different role.

The method template comes from quantitative genetics. Pedigree mixed models and PARSEC split trait variance into structured random effects, and PARSEC goes further by separating direct genetic effects from parental effects. PARSEC's lesson transfers directly: a component is only identifiable if the design breaks its confounding with other components. In genetics, sibling and parental genotypes do this. In our setting, donor-by-batch crossing does it. We do not use any genetic relatedness. We borrow the identification logic only.

The motivation comes from integration benchmarks and their known failure mode. Yi et al.'s benchmark of 16 deep integration methods showed that standard neighborhood metrics such as iLISI and kBET miss within-cell-type variation. They proposed the scIB-E metrics and a Corr-MSE loss that preserves cell-to-cell correlations. sysVI showed that strong KL regularization over-smooths biology and that adversarial training mixes unrelated populations. Both papers establish that overcorrection is real. Neither, as described in the file, separates donor variance from batch variance, and neither has an internal ground truth for what should be preserved. Our crossed-donor estimate provides that ground truth.

ResidPCA supplies the idea of first removing the cell-type component and then studying the residual. It also constrains our design. The file notes that subtracting cell type linearly can erase states that correlate with lineage. That is why our cell-type mean $\mu_{gk}$ is fit separately per type and why the batch term is allowed to vary by state bin, rather than one global regression being applied.

The case for counts comes from work on sparse data. scNBMF argues that negative-binomial likelihoods handle droplet data without log transforms. scMET isolates residual overdispersion after modeling the expected technical relationship between mean and variance. We reuse that principle, but for donor and batch components of expression rather than for highly variable methylation features. The file also lists imputation methods such as SERM and scRMD. Their documented risk of hallucinating correlations is exactly why we avoid imputing before decomposing.

Research Program

We will build three things.

The first is the estimator. It is a negative-binomial mixed model fit per cell type, using approximate restricted likelihood or variational inference, with a random-effect structure that matches the design. Fitting thousands of genes this way is expensive, so we will also implement a method-of-moments variant in the spirit of PARSEC's estimators for atlas-scale data.

The second is a semi-synthetic benchmark with known answers. We start from a single batch containing many donors and split each donor's cells across pseudo-batches. Each pseudo-batch then receives a known batch effect: gene-wise multiplicative distortions whose size depends on cell state, plus binomial thinning, which randomly deletes captured molecules to mimic lower capture efficiency and therefore more zeros. Because the donor variance is fixed by construction, we know exactly how much of it any method retains. We will vary thinning strength so the sparsity question is tested directly.

The third is the corrector. It estimates $v_{gb,z(c)}$ by the usual best linear unbiased prediction from the fitted model, subtracts it on the log-mean scale, and produces corrected expected expression and an embedding. Batch removal is then explicitly limited to the variance the model attributes to batch.

The real-data part uses datasets from the scIB-E benchmark and the cross-system settings in sysVI, such as single-cell versus single-nucleus and tissue versus organoid, wherever donors or matched samples appear in more than one batch. Where crossing is partial, we report how precisely each component is estimated, so that unidentifiable cases are visible rather than hidden.

Baselines

The naive reference point is the uncorrected data, analyzed with the same decomposition. Its donor variance is by definition fully retained. Its batch mixing, measured by iLISI and kBET, sets the floor. Any method must improve mixing to be worth anything, and our index measures what that improvement costs.

The strongest learned competitors are two. The first is sysVI, chosen because it was designed specifically to avoid over-smoothing under substantial batch effects. The second is the best Corr-MSE-trained model from Yi et al., chosen because it is the only method in the file explicitly optimized to keep within-cell-type structure. Each will be run with its authors' default settings and then given the same hyperparameter search budget as our corrector, on identical cells, genes, and compute. Selection uses a held-out batch-mixing target, so no method is tuned on our own index. The file reports no figures for these methods, so the comparison must be run fresh: no reported result found on Lacuna for sysVI or Corr-MSE scIB-E scores.

We also include BBKNN and iSMNN as fast neighbor-based methods. They modify the neighbor graph or match cells locally instead of training a generative model, which tests whether lighter correction alone preserves donor variance.

A pseudobulk baseline is the most serious statistical competitor. It sums counts per donor, batch, and cell type, then fits the same mixed model on those sums. It tests whether cell-level modeling adds anything.

Three ablations isolate the key ingredients. The first replaces the negative-binomial likelihood with a Gaussian model on log(x+1), which tests the sparsity claim. The second drops the state-dependent batch term, which tests whether batch effects interact with cell state, as Warnat et al.'s cross-platform analysis showed for microarrays (Warnat et al.). The third fits on a design with no crossed donors, to show that the estimate then collapses as expected.

Beating the baselines means two things. First, our corrector must retain significantly more of the true donor variance on the semi-synthetic benchmark, and more of the crossed-design donor variance on real data, while matching or exceeding sysVI's iLISI within its tuning budget. Second, our overcorrection index must rank methods differently from scIB-E's existing metrics in at least some settings. If its rankings simply agree with scIB-E's, it adds nothing.

Why not just…?

Why not just use scIB-E's intra-cell-type metrics? They measure whether structure inside a cell type survives integration, but they cannot say whether that structure is donor biology or leftover batch effect. A method that leaves residual batch variation would look good on them. The semi-synthetic benchmark settles this directly. We will check whether scIB-E scores correlate with the known true donor-variance retention across methods and thinning levels, and if they correlate strongly, the diagnostic half of this proposal is redundant.

Why not just pseudobulk and fit a standard mixed model? This is a real threat to the diagnostic. For estimating cell-type-level donor variance with many cells per donor, pseudobulk may be just as accurate and much cheaper. Its weakness is that it averages away the cell-to-cell component and any state-dependent batch effect, which reproduces the masking of heterogeneity that bulk sequencing causes. We expect cell-level modeling to win only in sparse or small cell types and for state-dependent effects. If pseudobulk matches it everywhere on the benchmark, we will adopt pseudobulk for the diagnostic and keep cell-level modeling only for the corrector.

Why not just include batch as a covariate in a conditional autoencoder? Conditional autoencoders already do this. The sysVI paper shows they over-smooth biology when KL regularization is strong. More fundamentally, they have no donor term, so nothing stops the batch covariate from absorbing donor variance when donors and batches are partly confounded. The comparison with sysVI, run at matched mixing, tests whether an explicit donor term matters.

Why not just run ResidPCA after integration? That treats a symptom. If integration has already removed donor variance, there is nothing left in the residual to recover, the same limitation the file notes for post-hoc latent methods such as scBatchProx. We will still report ResidPCA-derived state axes as a downstream readout of what each corrector preserves.

Decisive Checks

Three results decide the proposal, all on the semi-synthetic benchmark first and then on real crossed designs.

The first check is calibration. The count-based estimator should recover the planted donor variance within its confidence intervals across thinning levels. The Gaussian log(x+1) ablation should drift upward as sparsity increases, because extra zeros inflate apparent variance. If the two do not separate, the sparsity claim fails.

The second check is detection. The overcorrection index should expose methods that score well on iLISI but destroy donor variance, and it should disagree with scIB-E rankings in at least some settings.

The third check is the corrector itself. It should improve donor-variance retention over sysVI and Corr-MSE at equal batch mixing. If it improves retention only by mixing batches less, the corrector contributes nothing beyond the diagnostic, and we will say so.