Paired Deconvolution for Discordant Twins: Separating Lost Cells from Changed Cells

When one identical twin becomes obese and the other does not, any molecular difference between them cannot come from inherited DNA. That is why twin studies of fat biopsies are among the cleanest evidence that obesity itself remodels tissue. The best-known example reported a 47% loss of mitochondrial DNA and a sharp shutdown of branched-chain amino acid breakdown in the heavier co-twin's fat (Pietiläinen et al., 2008). The biopsies were ground up and measured in bulk, though, so the result leaves a basic question open. The fat cells themselves may have changed. Alternatively, the tissue may simply have filled with immune and stromal cells that carry fewer mitochondria, which dilutes the fat-cell signal. The standard fix, computationally splitting a bulk profile into cell types using a single-cell reference atlas (deconvolution), has a flaw that matters here. The reference comes from a different lab and platform, and that technical mismatch looks the same as a real change in cell makeup.

We propose a deconvolution model built around the twin pair rather than the individual. Both co-twins are profiled on the same platform and share the same genotype. Two of the main nuisances in deconvolution therefore hit both twins identically and can be absorbed into a term shared by the pair: the mismatch between reference and bulk platform, and genotype-driven differences in how each cell type expresses genes. With that term in place, what remains of the within-pair difference can be split into two parts: a shift in cell proportions and a cell-type-specific change in expression. A colleague could summarize it this way: the twin design already cancels genetics, and if the deconvolution model is written in paired form, it also cancels most of the reference batch problem, so we can ask which cells carry an acquired disease signal.

Core Claim

Unpaired deconvolution treats each sample's cell-type expression profile as the reference profile times a cohort-wide correction. Real individuals deviate from the reference for two reasons. One is platform bias. The other is their genotype: cell-type-specific genetic effects on expression are widespread, as CSeQTL's mapping from bulk data shows (CSeQTL). In an unrelated cohort, those per-person deviations end up as noise in the estimated proportions, or worse, as spurious cell-state changes. In a monozygotic pair the deviation is shared by both members. Our claim is narrow and testable. Modeling a pair-shared expression term should give calibrated tests for cell-type-specific acquired expression change under reference–bulk mismatch, and should reduce "leakage" of a change in one cell type onto another. Unpaired deconvolution followed by within-pair differencing should not achieve either. We do not claim the paired model improves proportion estimates much; simple differencing may already handle those, and our experiments are designed to check this.

Model

For gene $g$ in co-twin $i \in \{0,1\}$ of pair $j$, we model bulk counts as negative binomial with mean

$$ \mu_{jig} = s_{ji}\, \sum_{k} p_{jik}\, c_k\, x_{gk}\, b_{gk}\, e^{u_{jgk}}\, e^{\delta_{gk} d_{ji}}, $$

where the terms are defined as follows. The library size is $s_{ji}$ and $p_{jik}$ is the proportion of cell type $k$. The per-cell RNA content $c_k$ follows CSsingle's argument that cell size biases deconvolution (CSsingle). The reference expression is $x_{gk}$. The cohort-wide reference-to-bulk platform correction is $b_{gk}$, shared by all samples. The pair-shared deviation $u_{jgk}$ absorbs genotype and shared early environment and is shrunk toward zero with a learned variance. The indicator $d_{ji}$ equals 1 for the heavier co-twin. The acquired cell-type-specific log fold change $\delta_{gk}$ is the target, along with the proportion contrast $p_{j1k}-p_{j0k}$.

We work on raw counts rather than log-transformed values. CSeQTL and Count Bridges both show that transforms break the additive mixing of cell contributions (CSeQTL; Count Bridges). For microarray cohorts we use the analogous model on the linear intensity scale.

Identifiability is the central technical issue. The terms $b_{gk}$ and $u_{jgk}$ trade off against $x_{gk}$, so on their own they are not separately identified. The terms $\delta_{gk}$ and $p$ are identified only through within-pair contrasts that share $u$. We will state the conditions explicitly. One condition is that cell types differ in which genes they express. The other is that proportions vary within pairs. We will also report when they fail, for example when a gene is expressed in only one cell type and $\delta$ cannot be told apart from a proportion change. We keep $\delta$ sparse across cell types, which matches the biological question of whether the change sits in one lineage.

Research Program

The work has three parts.

The first is the model and its inference. We use expectation-maximization over proportions with penalized negative binomial updates for $\delta$, and we use multiple references in the spirit of SCDC to stabilize $x_{gk}$ (SCDC).

The second is a ground-truth benchmark of synthetic twins, which does not exist today. We build it from Onek1k, a blood single-cell cohort of about a million cells from over a thousand genotyped donors (used in Count Bridges). Splitting one donor's cells into two halves gives two "co-twins" with identical genomes. We plant known composition shifts and known cell-type-specific expression changes in one half, aggregate both halves to bulk, and deconvolve against a reference from a different study and platform, so the mismatch is real rather than simulated. Because Onek1k donors vary in genotype, we can measure directly how much of the error in unpaired methods comes from genetic expression differences. The Human Pancreas Multi-Study collection, which spans five sequencing technologies (scBatchProx), provides a second tissue and platform-swap setting.

The third is a reanalysis of the FinnTwin16 adipose biopsies from 14 BMI-discordant pairs (Pietiläinen et al., 2008; FinnTwin16), and of the related oxidative-phosphorylation twin data (Acquired obesity and poor physical fitness impair expression of genes of mitochondrial oxidative phosphorylation), if access is granted. The question is whether the mitochondrial and branched-chain amino acid signals localize to adipocyte $\delta$ or to a shift in proportions toward immune and stromal cells.

Baselines

The naive reference point is paired bulk differential expression with pair as a blocking factor. It detects acquired change but attributes it to no cell type. For attribution we compare against scMappR's cell-weighted fold changes, which redistribute bulk fold changes over cell types using estimated proportions and reference specificity (scMappR). On the Monaco PBMC benchmark it reports an average rho increase of 0.0471 in correlation with measured cell-type-specific differential expression over bulk fold changes.

The strongest proportion competitor is SCDC, which ensembles deconvolutions across several single-cell references weighted by reconstruction error (SCDC). It reports a Pearson correlation of 0.991 with ground truth on a 6:3:1 three-cell-line mixture. We run SCDC per sample, take within-pair differences, and test them with a paired test. This "deconvolve then difference" baseline is the main threat. CSsingle, run the same way, covers cell-size bias (CSsingle).

All methods get the same bulk data, the same reference set, and the same marker-gene selection. Every tunable method gets an equal hyperparameter search on a held-out set of synthetic pairs.

The key ablation is our model with $u_{jgk}$ removed, which treats co-twins as unrelated but keeps everything else. A second ablation removes $b_{gk}$.

Beating the baselines means the following. On null synthetic pairs, where only reference mismatch is present, our cell-type-specific tests must hold their nominal false-positive rate while differencing baselines do not. On planted pairs, we need higher power at matched false discovery and lower leakage of planted changes into non-target cell types. Matching SCDC on proportion contrasts is acceptable; that is not where we claim a gain.

Why not just…?

Why not just deconvolve each twin with SCDC and subtract? For proportions this may work nearly as well, and we expect it to. Differencing does not produce cell-type-specific expression change. It also cannot tell whether a gene's fold change reflects proportion or state, because the per-person genetic deviation in expression enters nonlinearly through the mixture rather than cancelling in a difference. The ablation with and without $u$, run across Onek1k donors grouped by genetic effect strength, settles this. If removing $u$ costs nothing, the pair-shared term is unnecessary and we will say so.

Why not just run paired differential expression with estimated proportions as covariates? This is the common practice. CSeQTL showed that ordinary regression on transformed bulk data leaks effects from abundant cell types into rare ones and inflates false positives (CSeQTL). Our leakage metric on planted pairs measures this directly, and we include this regression as an additional baseline.

Why not just collect single-nucleus data from new twin biopsies? That is the eventual gold standard. Discordant twin cohorts are rare and archived, however. The one single-cell twin study in the file profiled a single pair (CVID atlas), and droplet protocols capture adipocytes poorly, which is a protocol bias SCDC was built to address. Our method makes existing bulk twin archives interpretable now. Its predictions on FinnTwin16 can then be checked against a small follow-up nucleus study, which would be the external validation.

Proposed Experiments

The first decisive check is null calibration. On Onek1k synthetic twins with no planted changes, deconvolved against a reference from another platform, we compare the quantile–quantile behavior of $\delta$ tests across all methods. The second is attribution under planted changes. We vary effect size, target-cell abundance down to rare types (scMappR reports detection in populations under 5% of cells), and the strength of genetic effects. We report power, false discovery, and leakage. The third is a positive control on the HbA1c association in 77 pancreatic islet donors, where SCDC recovered the inverse beta-cell relationship at p = 0.0019 (SCDC). It is unpaired, so it only checks that our proportion component does not degrade.

The final test is the FinnTwin16 reanalysis. With 14 pairs, only large effects are detectable, and we will report intervals rather than claims. The outcome is informative either way. If the mitochondrial loss sits in the adipocyte term, the "brownout" is intracellular. If it is absorbed by proportion shifts toward immune and stromal cells, the classic finding is partly an artifact of the tissue's cell makeup. If weight-concordant control pairs are available, they serve as a within-cohort null that should show neither.