Identical twins are used in medical research because they share an inherited genome. When one twin develops a disease or gains weight and the other does not, the difference is attributed to environment rather than genes. That attribution rests on an assumption almost never checked. Mutations that arise after the embryo splits can leave one twin with patches of cells carrying DNA changes the co-twin lacks. Standard confirmation pipelines miss these. Those pipelines sequence the exome and then re-check candidate variants with Sanger sequencing (a slow but accurate method that reads one DNA fragment at a time), and Sanger cannot see a variant carried by fewer than roughly 15–20% of cells in a bulk sample.
This proposal turns the twin design into a measurement tool. Single-cell RNA sequencing reads the active genes of thousands of individual cells, and those reads also carry the cell's DNA sequence at expressed positions. Within a twin pair, the co-twin's cells of the same type are a perfect germline control: any variant both twins share was inherited or arose before twinning, and any variant one twin has and the other lacks arose afterward. We propose to call these twin-discordant mutations per cell type from single-cell data and confirm them with DNA assays, the twins' parents, and deep targeted sequencing. We then ask whether the mutated cells, compared with unmutated cells of the same type in the same person, explain any of the gene-expression differences between the twins.
Core Claim
The co-twin is a stronger reference for single-cell somatic variant calling than either an unrelated panel of normals or a population reference panel, for two reasons. First, the co-twin removes private germline variants and person-specific artifacts that population resources cannot remove. Second, it resolves a specific failure of existing callers: telling an early somatic mutation present in about half of cells apart from an ordinary heterozygous germline variant.
Once such variants are confirmed, each cell's genotype at the mutated site becomes a label. That label lets us compare mutant and wild-type cells within one person and one cell type. This replaces the artificial overexpression assays used in current validation pipelines with an endogenous readout. It also lets us test, rather than assume, the "genetically identical" premise of discordant-twin expression studies.
Why This Is Not Already Done
Three bodies of work each cover part of this.
The first is discordant-twin molecular studies, which motivate the question. Global Transcript Profiles of Fat in Monozygotic Twins Discordant for BMI compared adipose microarrays across 14 body-mass-discordant pairs. They sequenced mitochondrial DNA to exclude it, but not nuclear somatic variation. Garske et al. profiled chromatin and expression in cultured preadipocytes from discordant pairs and found 52 differentially expressed genes in remodeled regions, again assuming nuclear identity. The closest work is the CVID single-cell atlas. It profiled one discordant twin pair with single-cell RNA, methylation, and chromatin assays and used bulk whole-genome sequencing to rule out causal germline variants. It never called variants from the single-cell reads, so lineage-restricted mosaicism in the affected twin's B cells was not examined. It also validated in 10 unrelated patients versus 10 controls, which drops the paired structure.
The second is single-cell variant callers, which supply our starting method. SComatic splits reads by cell type and models sequencing error with a beta-binomial background (a count model whose error rate is allowed to vary from site to site). It removes recurrent artifacts with a panel of unrelated normal samples and requires support from at least three reads in two cells of one type. Monopogen uses linkage patterns from external population panels to recover germline variants. By its own account it struggles to separate early truncal somatic mutations from germline variants near 50% allele frequency, and it has limited power for private mutations. Neither has a person-matched germline control, and neither confirms calls with DNA.
The third is Sanger-plus-pedigree validation studies, which constrain the design. The ABCD1 study and the DEPDC5 epilepsy series show the standard cascade: exome discovery, Sanger confirmation, parental testing, and in vitro assays. That cascade has two weaknesses. It is blind to low-level mosaicism. Its functional step is artificial: the ABCD1 authors saw mutant mRNA and protein rise in 293T kidney cells and acknowledge that the cell line is not the target tissue.
Research Program
The calling method compares three explanations at each candidate site within each annotated cell type $c$. Let $a_A, d_A$ be alternate-allele reads and total reads in twin A's cells of type $c$, and $a_B, d_B$ the same in twin B. The three explanations are:
- artifact: both twins at the site-specific error rate;
- shared variant: both twins carry the allele;
- discordant variant: twin A carries it at fraction $f$ while twin B shows only error.
The discordance score is
$$S_c = \log \frac{\mathrm{BB}(a_A \mid d_A, f)\,\mathrm{BB}(a_B \mid d_B, \epsilon_s)}{\max\{\mathrm{BB}(a_A \mid d_A, \epsilon_s)\,\mathrm{BB}(a_B \mid d_B, \epsilon_s),\ \mathrm{BB}(a_A \mid d_A, f')\,\mathrm{BB}(a_B \mid d_B, f')\}}$$
In this score, $\mathrm{BB}$ is a beta-binomial likelihood, $\epsilon_s$ is SComatic's site-specific background error, $f$ is twin A's allele fraction, and $f'$ is a shared allele fraction. Both fractions are fit by maximum likelihood.
The key requirement is evidence of absence. Twin B must have enough reads at the site in the same cell type for its zero alternate reads to count as informative. Low-coverage sites are left unresolved rather than called. This is what an unrelated panel of normals cannot provide: it flags only sites that are recurrent across people. The same comparison handles Monopogen's ambiguity. A variant near 50% in twin A and absent in a well-covered twin B is post-twinning somatic, whatever its allele fraction.
Every discordant call then enters a DNA cascade that adapts the validation pipelines above. First, deep targeted amplicon sequencing of DNA from sorted cell populations of both twins. This assay can measure allele fractions far below Sanger's limit. Sanger is reserved for high-fraction calls. Second, parental genotyping. A variant found in a parent but in only one twin indicates an error or loss of heterozygosity, not a new mutation, and this step catches that. Third, a DNA check that removes RNA editing: a twin-specific inflammatory state could plausibly change RNA editing at a site and mimic a discordant mutation, and only DNA can separate the two.
For confirmed variants, we label each cell as mutant, wild-type, or uncovered. There is a trap here. A cell can be genotyped only if it expresses the gene at the mutated site, so "covered" cells are already selected for high expression of that gene. A naive comparison of mutant cells against all other cells therefore confounds genotype with expression level. We compare mutant against wild-type cells only among covered cells. We also match on total counts and on the site gene's expression, and we report the uncovered fraction. Tests use a within-person permutation of genotype labels among covered cells, which keeps the cell-type and sequencing structure intact.
At the pair level, intra-pair differences per cell type are estimated by paired pseudobulk (summing each twin's reads within a cell type and comparing the twins). We then ask what fraction of a cell type's twin-discordant expression program is reproduced by the within-twin mutant-versus-wild-type contrast.
The primary cohort is body-mass-discordant pairs recruited with the Garske criterion of at least 3 kg/m² difference. Sample size will realistically be on the order of the 14 pairs in Naukkarinen et al. For each twin we collect fresh subcutaneous adipose stromal-vascular cells and blood for single-cell RNA sequencing, deep bulk whole-genome sequencing of blood and adipose DNA, and DNA from both parents where available. A secondary analysis reruns the pipeline on the CVID twin pair's B-cell data, if access can be obtained, where lineage-restricted clones are biologically plausible.
Baselines
The naive reference is the classical assumption that discordant identical twins carry no relevant genetic difference. In practice this means reporting paired pseudobulk expression differences with no variant analysis, as the bulk twin studies did. This baseline costs nothing and will be run on the identical cohort. It defines the discordance we try to explain.
The strongest existing caller is SComatic run with its default panel of normals and calling thresholds on the same reads, cell-type annotations, and coverage filters. The reported figures are precision 0.67–0.87 and F1 0.6–0.7 in SComatic, against 0.2–0.4 for the next-best caller. Beating it means higher precision at matched recall on the spike-in and matched-sequencing benchmarks below, plus a higher amplicon-confirmation rate on real twin calls. Monopogen will be run with its default population panel. Its reported 95% germline genotyping accuracy against matched whole-genome sequencing (Monopogen) is the reference for the germline layer. The specific test is whether the twin comparison reclassifies near-50% calls that Monopogen labels germline, and whether amplicon sequencing confirms those reclassifications.
For expression, the reference methods come from the nested single-cell differential expression benchmark. It reports AUPRC (area under the precision-recall curve) of 0.93 for pseudobulk DESeq2 in balanced designs. Under unbalanced conditions the permutation test reached 0.94 while DESeq2 degraded. We will run paired pseudobulk DESeq2 with pair as a blocking factor, and a paired permutation test, with identical gene filters.
Two ablations isolate the key ingredient. The first adds the co-twin's reads to SComatic's normal panel rather than using the coverage-aware comparison. The second drops the evidence-of-absence coverage requirement. A third ablation, for the expression step, compares mutant cells against all cells rather than coverage-matched wild-type cells, to show the size of the selection bias.
Why not just…?
Why not just deep-sequence bulk DNA from both twins and compare? For clones that dominate a tissue this is a real threat, and it is included in our design as an independent caller and as validation. It falls short in two ways. A mutation restricted to a minor cell type, such as adipose macrophages, is diluted below detection in bulk tissue. And bulk DNA cannot say which cells carry the variant, so it cannot support the within-person mutant-versus-wild-type expression contrast that is the point of the study. We will report the overlap between bulk-detected and single-cell-detected discordant variants, stratified by the abundance of the carrying cell type.
Why not just add the co-twin to SComatic's panel of normals? This is the simplest form of the idea and is the first ablation above. A panel of normals removes a site whenever it appears anywhere in the panel. It does not separate "absent despite good coverage" from "unobserved," and it cannot label shared early variants as somatic rather than germline. If the ablation matches the full method on the spike-in benchmarks, we will adopt it. The contribution then reduces to the validation cascade and the genotype-stratified expression analysis, and we would say so.
Why not just run paired pseudobulk expression analysis and skip variants? It will find twin-discordant genes, and it is our naive baseline. It cannot tell whether a difference reflects acquired environment or acquired mutation. That distinction is the unchecked premise behind the BMI-twin findings, and resolving it is the question this study asks.
Proposed Experiments
The first benchmark is a pseudo-twin calling test. Multiple heart regions from one donor share a germline exactly. We pair such samples from the matched-genome heart data used by Monopogen, and from the Human Heart Cell Atlas used by SComatic, as stand-in twins. Into one sample we inject alternate-allele reads at known sites. The injections cover a fraction of cells of a chosen type and span a grid of allele fractions, including near-50% "truncal" fractions. We score precision and recall against SComatic, Monopogen, and the ablations.
The second benchmark uses real mutations. In the cutaneous squamous cell carcinoma data with matched exome sequencing analyzed by SComatic, tumor and matched normal skin from the same person form a germline-matched pair. This checks the comparison against DNA-confirmed truth.
The decisive checks in the twin cohort follow from these. The method is useful only if it does three things:
- yields twin-discordant calls whose amplicon-confirmation rate exceeds SComatic's default calls on the same pairs;
- reclassifies some near-50% variants as post-twinning, with DNA confirmation;
- keeps the coverage-matched mutant-versus-wild-type contrast calibrated under permuted labels.
The biological result is informative either way. If confirmed mosaic variants in a cell type reproduce part of that cell type's twin-discordant expression program, the "genetically identical" premise needs revision for that tissue. If no confirmed variant explains any discordance, the environmental interpretation of studies like Naukkarinen et al. and Garske et al. gains, for the first time, a direct genetic check at single-cell resolution. No reported result on Lacuna was found for the rate of twin-discordant somatic variants in adipose tissue, so we make no numerical prediction for it.
References
- Global Transcript Profiles of Fat in Monozygotic Twins Discordant for BMI: Pathways behind Acquired Obesity
- Increased body mass index is linked to systemic inflammation through altered chromatin co-accessibility in human preadipocytes
- Single-cell Atlas of common variable immunodeficiency shows germinal center-associated epigenetic dysregulation in B-cell responses
- De novo detection of somatic mutations in high-throughput single-cell profiling data sets
- Single-nucleotide variant calling in single-cell sequencing data with Monopogen
- A case of adrenomyeloneuropathy caused by a novel point mutation in the ABCD1 gene and functional verification
- Case Report: Unraveling clinical heterogeneity in DEPDC5-related epilepsy: a genotype–phenotype correlation study in eight pediatric cases
- Single-cell differential expression analysis between conditions within nested settings