Genetics of the Fat-Cell Transition: Stage-Resolved Heritability for Fat Distribution

Fat tissue is not one kind of cell. It holds mature fat cells, a large and varied pool of precursor cells, blood vessel cells and immune cells, and the precursors must pass through intermediate stages before they fill with lipid. Genetic studies of body shape usually measure this tissue as a single ground-up average (bulk RNA sequencing). Recent single-cell work that ties obesity genetics to fat tissue has looked only at mature fat cells. So if some of the unexplained ("missing") heritability of fat distribution sits in variants that act only while precursors are becoming fat cells, or in variants that change how many precursors commit to that path, current tools cannot see it.

We propose to reconstruct differentiation-stage-resolved gene expression for each individual from existing genotyped bulk fat samples, using a human fat-tissue single-cell atlas as the reference. We will correct explicitly for the fact that mature fat cells and precursors produce very different amounts of RNA. We then ask how much heritability of waist-to-hip ratio is explained by genetic effects confined to transitional precursor states, over and above what bulk tissue, mature fat cells, and simple marker-gene enrichment already explain. In two sentences: stage-specific genetic effects during fat-cell formation are averaged away in bulk tissue and skipped by adipocyte-only studies. We recover them by genetically aware, activity-corrected unmixing of bulk fat and test whether they carry heritability that nothing simpler captures.

Core Claim

The claim is that part of the heritability of fat distribution acts through one of two routes. The first is expression differences that exist only in committed and transitional preadipocytes. The second is genetic control of the ratio between precursors that commit to becoming fat cells and those that become structural stroma. We also claim that recovering either route from bulk tissue requires separating cell proportion from per-cell RNA output. Without that separation, the enlarged fat cells of heavier people inflate apparent adipocyte fractions and create false genetic signals.

Three lines of prior work motivate this claim. The SREBF1 trans-effects study paired single-nucleus data from 7 individuals with 391,701 UK Biobank participants. It found that subcutaneous adipocyte marker genes carry a 2.21-fold enrichment for abdominal-obesity heritability, and that SREBF1 risk variants regulate 146 adipocyte genes in trans. That study examined only mature adipocytes and says nothing about precursor stages. The integrated white adipose tissue map combined 401,320 cells from 103 samples. It resolved 17 precursor clusters in subcutaneous fat, falling along two trajectories: committed preadipocytes versus anti-fibrotic structural stroma. It also deconvolved 864 bulk samples, but used static signatures, no genotypes, and no heritability model. The mechanistic literature says the decisive checkpoints sit at the precursor stage. ACAT1/SOAT1 must esterify cholesterol before PPARγ can be induced. Annexin A1 brakes commitment by triggering SMAD4 degradation. NF-Ya gains stage-dependent binding to the Fasn promoter while its protein level stays constant. circITGB1 tilts precursors toward proliferation rather than differentiation. These studies supply prespecified genes whose genetic effects our hypothesis predicts should be stage-confined.

Research Program

The method comes mainly from the SPLITR brain deconvolution work. SPLITR unmixed 3,387 post-mortem brain samples while separating cell abundance from per-cell transcriptional activity (neurons make about four times more mRNA than glia). It found 1,182 cell-type eGenes, meaning genes with a genetic effect on expression, that bulk analysis missed. It also ran a GWAS on cell fractions. We move this logic to fat, where the activity gap is plausibly larger and, unlike in brain, covaries with the trait itself, since adipocyte size tracks BMI.

We replace discrete cell types with positions along the precursor-to-adipocyte trajectory from the adipose map. For individual $i$, the bulk count of gene $g$ is modeled as

$$E[y_{gi}] = s_i \sum_{k} \pi_{ik}\, a_{k}(x_i)\, \mu_{gk}\, \exp(\beta_{gk}\, G_{gi}),$$

where $s_i$ is sequencing depth, $\pi_{ik}$ is the fraction of cells in stage $k$, $a_k(x_i)$ is the per-cell RNA output of stage $k$ given covariates $x_i$ (BMI, sex, age), $\mu_{gk}$ is the reference expression profile from the atlas, $G_{gi}$ is the individual's genotype at a nearby variant, and $\beta_{gk}$ is the stage-specific genetic effect on expression. The stage set $k$ has three kinds of members: mature adipocytes, transitional preadipocytes (early pseudotime bins along the commitment trajectory), and structural stroma. Immune and vascular compartments enter as nuisance components. A negative binomial likelihood handles count noise. We deliberately use a linear-mixture model rather than a deep generative one; the Count Bridges model is reserved for a robustness check. This keeps the key ingredient, the activity term $a_k$, isolated and interpretable.

We fit the model to genotyped bulk subcutaneous fat from GTEx. The outputs are stage-resolved eQTL estimates, stage-resolved expression-prediction weights, and per-person stage fractions. Those fractions feed two analyses. The first is a GWAS of the commitment ratio, the fraction of committed preadipocytes relative to structural stroma. The second is a set of stage-specific transcriptome-wide association tests, which impute each person's genetically predicted gene expression into a large cohort and test it against the trait, here UK Biobank waist-to-hip ratio stratified by sex. Heritability is then partitioned across stages with Gene Score Regression, which corrects the association statistics for correlation between neighboring genes. This removes the linkage artifacts that GSR showed contaminate bulk TWAS, where gene-score/trait correlation reached 0.76 in schizophrenia. Every test also reports direction of effect, which scPrediXcan omits.

To tie human stages to known mechanism, we map the shACAT1 3T3-L1 differentiation time course onto human pseudotime. We then check whether genes disrupted early by ACAT1 knockdown are over-represented among transitional-stage eGenes.

Baselines

The naive reference point is bulk subcutaneous-adipose TWAS using GTEx weights. Its GSR heritability estimate and its count of independent significant loci for waist-to-hip ratio define the floor. We found no reported result on Lacuna for bulk adipose TWAS on waist-to-hip ratio, so we will compute it ourselves on the same GWAS summary statistics and LD reference.

The second baseline is mature-adipocyte-only analysis in the style of the SREBF1 study. Its reported number is a 2.21-fold enrichment for subcutaneous adipocyte markers. We will reproduce that enrichment with the atlas's adipocyte markers under our pipeline, then ask whether transitional-stage genes add heritability conditional on adipocyte genes.

The strongest existing method is scPrediXcan. Its sequence-based expression predictor needs only per-stage mean expression, not per-stage genotypes, so it could in principle be trained on atlas pseudotime bins. It reports 222 candidate genes across 108 LD blocks for type 2 diabetes, against 111 for bulk TWAS and 12 for pseudobulk TWAS. We will train its linearized version on the same atlas stage means, apply it to the same waist-to-hip statistics, and give it the same compute and hyperparameter search as our weights. We count as beating it only if our stage weights both explain more GSR-partitioned heritability in transitional stages and replicate more often in held-out data.

The key ablation removes the activity term by fixing $a_k$ to one. If fraction estimates and transitional-stage heritability do not change, the activity correction is not doing work.

Why not just…?

Why not just run stratified heritability enrichment on the atlas's precursor marker genes? This is the cheapest serious alternative and a real threat for the enrichment question alone. If transitional-stage marker sets already show the enrichment, our stage-resolved eQTLs may add nothing to the headline number. We therefore make marker-set enrichment a conditioning baseline. Our claim survives only if stage-specific genetic weights explain heritability beyond marker membership, and if they name specific genes with directions of effect that marker sets cannot supply. If they do not, we will report that transitional-stage heritability is real but adequately captured by markers.

Why not just fit cell-fraction-by-genotype interaction eQTLs on bulk GTEx using fractions from standard deconvolution? This is essentially our model without the activity term and with fractions estimated separately, so the ablation above tests it directly. SPLITR's immunostaining validation showed that standard CIBERSORT distorted fractions and missed rare populations when activity and phenotype shifts were ignored. In fat, uncorrected fractions should correlate spuriously with BMI through adipocyte hypertrophy. We will measure that correlation in both models. If the simple version matches ours on simulation recovery and real heritability, we will adopt it.

Why not just train a sequence-based predictor on atlas stages and skip deconvolution? A sequence model learns how much DNA sequence explains average expression differences between genes. It cannot learn how inherited variants shift expression between people within a stage, especially at trans-acting hubs like SREBF1. The head-to-head comparison with scPrediXcan, including replication in RYSA (SREBF1 study), will settle whether individual-level stage eQTLs matter here.

Proposed Experiments

The first experiment is a calibration simulation. Pseudobulk samples built from held-out atlas donors are paired with GTEx genotypes, and known eQTLs confined to transitional stages are planted in them. We measure how well each method recovers the planted eQTLs, using area under the precision-recall curve. For reference, GSR reports 0.925 versus 0.260 for PASCAL in its own simulation setting. We also measure false positives when BMI-linked activity shifts are simulated with no genetic effect. The activity ablation is expected to fail specifically in that null.

The main experiment partitions waist-to-hip heritability across stages, conditional on adipocyte markers. The decisive check is whether conditional heritability for transitional stages is significantly above zero in females, where the SREBF1 study found its signal strongest.

The third experiment is the commitment-ratio GWAS. The decisive check is whether any genome-wide locus replicates when ratios are re-estimated from the atlas's own single-cell donors, where fractions can be counted directly.

The fourth is a prespecified mechanistic test. ACAT1/SOAT1, ANXA1, SMAD4, NFYA and SREBF1 targets should show stronger genetic effects on expression in transitional stages than in mature adipocytes, and early ACAT1-knockdown genes should be enriched among transitional eGenes.

A null result on the main experiment, with calibration passing, would be informative in its own right. It would indicate that adipose-mediated missing heritability of fat distribution sits in mature adipocytes, and that the precursor biology, however mechanistically rich, is not where common genetic variation acts.