Haplotype Effects on Live Volumetric Cell Dynamics Across Ancestries

When a genetic variant influences disease, it usually acts long before the clinic sees anything: it changes how a cell's regulatory genes switch on, how fast its organelles reorganize, how quickly it changes shape under stress. Two measurement barriers have kept genetics blind to these intermediate steps. First, the standard molecular assay (RNA sequencing) requires destroying the cells, so we can never watch the same cells' program unfold — we infer dynamics from parallel cultures sampled at different times. Second, the imaging that does capture living cells in three dimensions has almost never been pointed at genetically diverse human cohorts, so we know a great deal about what cells do on average and almost nothing about how DNA sequence variation makes individual donors' cells behave differently. Meanwhile, the genetics side has its own blind spot: association studies test single DNA letters one at a time, even though causal effects often come from combinations of variants inherited together on one chromosome — a haplotype — and even though those single-letter tests are known to transfer poorly from European to African, Asian, and admixed populations.

This proposal builds a single study that closes all of these gaps at once, and in doing so tests a concrete scientific claim: that many genetic effects on cellular behavior are invisible to both static and single-variant analysis, and that haplotype-based tests, applied to trajectories of three-dimensional cell phenotypes rather than to one-time gene counts, recover effects that replicate across ancestries where single-variant tests fail. The core deliverable is a method — call it a dynamic haplotype QTL model — plus the multi-donor volumetric imaging dataset needed to run it.

Core Claim

The claim has three parts. First, cellular phenotypes measured as time courses of volumetric snapshots (whole-cell 3D shape, dry-mass distribution, organelle architecture, tracked over hours) are heritable intermediate traits in which genetic effects are phase-specific — active during early response and silent later, or vice versa. This is expected because the one existing dynamic study of this kind found that 23% of genes with expression QTLs showed division-state-dependent effects, with disease variants like RPS26 acting mainly during early activation (Division state reveals hidden genetic regulation during T cell activation and identifies immune disease–linked gene programmes). Second, a substantial share of those effects are carried by multi-variant haplotype blocks rather than any single marker, mirroring what the eQTLHap framework demonstrated for static expression: 22 validated associations in GEUVADIS and GTEx that single-SNP models could not detect (A novel haplotype-based eQTL approach identifies genetic associations not detected through conventional SNP-based methods). Third — and this is the part with the widest reach — haplotype-defined effects should transfer across ancestries better than single-tag-SNP effects, because the haplotype is the physical unit of inheritance and tags the ancestral chromosome itself, whereas a tag SNP's correlation with the causal variant is reshuffled independently in each population. The literature sets up this tension but never resolves it: fine-mapping orthodoxy says multi-ancestry cohorts should break linkage disequilibrium down to single causal nucleotides, while haplotype evidence from eQTLHap, AlleleDB, and SparsePainter says the causal unit is often the compound block. Nobody has asked which unit survives transfer across ancestries for dynamic cellular traits.

Research Program

The study design uses lymphoblastoid cell lines (LCLs) from 1000 Genomes donors — the same resource behind GEUVADIS expression data and AlleleDB's allele-specific catalog of 63,541 expression and 6,136 binding variants across 382 diverse individuals (A uniform survey of allele-specific binding and expression over 1000-Genomes-Project individuals). LCLs are the right substrate because they are immortalized (parallel cultures can be harvested destructively at each time point without donor re-recruitment), genotypes and statistical phase already exist at scale, and the donor panel spans the ancestry breadth the third claim requires. We select roughly 300 donors balanced across European, African, East Asian, and admixed backgrounds.

The imaging arm applies an existing snapshot volumetric platform rather than building new hardware. The template is Mantis, which combines oblique light-sheet fluorescence with label-free quantitative phase imaging and already demonstrated 4D profiling of 20 cell lines at 15-minute intervals over 7.5-hour windows at transverse phase resolution below 400 nm, designed for up to a million cells per experiment (Mantis: high-throughput 4D imaging and analysis of the molecular and physical architecture of cells). Mantis itself was run on homogeneous cell lines with no genotypes; we keep its instrument and analysis pipeline and change only the cohort, imaging donor-by-donor across a stimulus time course (immune activation, chosen to match the timescales where the T-cell study found phase-specific genetic effects). Snapshot acquisition matters here specifically because it removes the temporal skew that mechanical scanning introduces across z-planes — different depths of a moving cell would otherwise correspond to different biological instants, corrupting exactly the trajectory features we want to associate with genotype (Shimizu et al.; Carlton et al.). In parallel, matched replicate cultures are RNA-sequenced at the same time points, following the established pattern of longitudinal profiling studies (Longitudinal analysis of gene expression profiles using functional mixed-effects models), so that morphological and transcriptomic trajectories can be linked donor by donor.

The statistical core is a model that maps haplotype dosages onto phenotype trajectories without the degrees-of-freedom explosion that naively multiplying haplotype states by time points would cause — the central tension identified in the gap analysis. For donor $i$ with haplotype dosage vector $h_i^{(b)}$ at LD block $b$ (each entry counting 0, 1, or 2 copies of a given ancestral configuration, as in eQTLHap's Bag-of-Haplotypes encoding), we model a volumetric trait trajectory $y_i(t)$ as

$$y_i(t) \;=\; \sum_b h_i^{(b)\top} \beta_b(t) + \epsilon_i(t), \qquad \beta_b(t) = \sum_{k=1}^{K} \gamma_{bk}\, \phi_k(t),$$

where $\phi_k(t)$ are a small set of temporal basis functions (e.g., early-response and late-remodeling splines) and $\gamma_{bk}$ are penalized coefficients. A sparse group-lasso penalty over (block, haplotype) × basis groups selects sparse blocks and, within a block, sparse temporal phases, which is what keeps the multiple-testing burden tractable. The same machinery runs on transcriptomic trajectories, giving directly comparable haplotype-dynamic expression QTLs. Two further components address known failure modes: local ancestry inferred with a PBWT-based painter (Sparse haplotype-based fine-scale local ancestry inference at scale reveals recent selection on immune responses) enters as a covariate so that admixed donors' mosaic chromosomes do not confound haplotype effects; and within-donor allele-specific expression, measured at heterozygous sites with the beta-binomial approach of AlleleDB, serves as an ancestry-free internal check on any cis-regulatory haplotype effect we call, since allele imbalance within one individual cannot be explained by population stratification.

Proposed Experiments

The first experiment is a replication sanity check: we reimplement eQTLHap on GEUVADIS static expression and confirm recovery of its reported 22 single-SNP-invisible associations, and reproduce its reported robustness result that phasing switch errors reduce sensitivity by less than 4%. This verifies our phasing and block-partitioning stack before any new data are collected; if we cannot reproduce the 22, the pipeline is wrong and nothing downstream proceeds.

The second experiment is the discovery study on the new imaging cohort. For each volumetric and transcriptomic trait trajectory, we run three association models on identical data: single-SNP dynamic (the design of the T-cell study), haplotype-static (eQTLHap applied to a terminal time point), and our haplotype-dynamic model. Primary metrics are the count of associations passing a matched false-discovery threshold, the fraction validated by allele-specific expression within heterozygous donors, and trajectory-prediction $R^2$ on held-out donors. The T-cell study's CellDivider model, which predicts division state from single-cell transcriptomes at Pearson $r = 0.881$ and MAE $0.428$, provides the calibration target for how well a latent-time model should track state — we retrain an equivalent model on our volumetric features and expect comparable or better trajectory inference since imaging time is observed directly rather than inferred.

The third experiment is the ancestry transfer test, and it is the decisive one for the core claim. Effects discovered in European donors are tested for replication in held-out African, East Asian, and admixed donors, comparing three predictor representations at matched discovery stringency: the tag SNP, the fine-mapped best single variant, and the haplotype block. The claim predicts a strictly ordered replication gradient — haplotype ≥ fine-mapped SNP > tag SNP — and specifically that some haplotype effects replicate where no single variant in the block does. A null result here (haplotypes transfer no better than tag SNPs) would itself settle the Gap-5 contradiction in favor of the fine-mapping orthodoxy, which is a publishable outcome either way.

The fourth experiment is a simulation calibrated to the cohort: we inject known haplotype effects of varying phase-specificity into simulated diverse-cohort genotypes and trajectories, measuring power under realistic phasing error rates, admixture proportions, and donor counts. This is where we quantify whether 300 donors suffice, using the eQTLHap switch-error simulations as the starting point for the error model.

Baselines

The naive reference point is the field's implicit default: a static, single-SNP association test on the terminal time point with global ancestry adjustment via top principal components. This is what nearly every existing eQTL catalog effectively is, and it is the configuration that eQTLHap showed misses real effects and that Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing showed misclassifies non-European patients. We run it on our own data as the floor; any effect it finds is a lower bound on what the richer models should recover.

The strongest competitor is the dynamic single-SNP design of the T-cell activation study (Division state reveals hidden genetic regulation during T cell activation and identifies immune disease–linked gene programmes), which is the only prior work that maps genetic effects as a function of a dynamically tracked cell state, and which found 23% of eQTL genes to be state-dependent — precisely the phenomenon our first claim extends. We implement it faithfully, including its latent-state regression, and run it on the same donors, the same trajectories, the same covariates, and the same compute and tuning budget as our model. Beating it means two things concretely: recovering strictly more cross-validated, ASE-validated associations at matched FDR, and finding at least a defensible set of loci where the single-SNP model detects nothing and the haplotype model detects an effect that replicates in held-out donors. If the single-SNP dynamic model finds everything we find, the haplotype ingredient earns nothing and the proposal's central bet has failed.

The key ablation replaces the haplotype dosage encoding with unphased single-SNP dosages inside our own trajectory model, holding the temporal basis, penalization, and covariates fixed. This isolates the one ingredient the proposal actually adds beyond prior dynamic-QTL work; the difference between this ablation and the full model is the marginal value of phase-aware prediction, and it is the number a skeptical reader should ask for first. A second ablation collapses trajectories to their terminal time point, isolating the temporal ingredient on top of the haplotype ingredient.

Why not just…?

Why not just use standard 2D Cell Painting or confocal z-stacks? For slow morphological traits this objection has real force: the 9-week radiation study successfully predicted five of nine nuclear morphology features from 2D Cell Painting measurements (Interpretable transcriptome-to-phenotype modeling of cell-painting nuclear morphology features from RNA-seq under low-dose radiation exposure), so if our traits were nuclear size and texture, 2D would plausibly suffice. We concede this for those traits and include them deliberately. The traits where snapshot volumetric imaging is not replaceable are the fast 3D ones — organelle redistribution, volumetric dry-mass dynamics, membrane deformation during activation — which Mantis showed carry physical signatures 2D imaging cannot see, and which mechanical scanning distorts through plane-to-plane temporal skew and phototoxicity (Mantis; Shimizu et al.). The experiment that settles it is already in the design: we project every volumetric trait to its best 2D summary and compare genetic effect recovery, 3D versus 2D-projected, on identical donors. If the 2D projections recover all the associations, the imaging arm was unnecessary and we will say so.

Why not just run standard single-SNP eQTL mapping? Because the evidence that it misses real biology is now direct rather than hypothetical: 22 replicated associations invisible to single-SNP tests in GEUVADIS and GTEx (eQTLHap), the PIN1 promoter where two mutations more than double transcriptional output only when both sit on the same chromosome, and the KLC1 haplotype whose cataract risk is undetectable marker-by-marker. The standard objection — that phasing errors and added degrees of freedom destroy the power gain — is quantified and weak: switch errors cost under 4% sensitivity in eQTLHap's simulations, and our group-structured penalization is designed precisely for the dimensionality problem. Still, we do not argue this in the abstract; the single-SNP dynamic model is our strongest baseline, run head-to-head, and the ablation against unphased dosages measures the marginal value exactly.

Why not just profile one time point in a much larger cohort? The strongest existing evidence against this is that genetic effects are phase-specific: 23% of eQTL genes act differently across division states, and RPS26-type disease variants act early and go silent later (Division state reveals hidden genetic regulation during T cell activation); myocardial infarction transcriptomes invert from inflammatory to metabolic programs between acute and chronic phases (Zhihua Liu et al.); and chromatin opens days before transcription rises (Ruikang Li et al.). A single snapshot sampled at the wrong phase averages these away. The fair version of the objection is a power trade-off — one time point buys several-fold more donors — and our simulation experiment is built to answer it: we compare a 300-donor trajectory design against a simulated larger single-time-point cohort at equal budget, and if the static design wins on validated discoveries, the temporal arm should be dropped.

Why not just adjust for the top 10 genetic principal components? Global PCs demonstrably miss fine-scale structure: in UK Biobank, SNP principal components predict birth location at a median error of 77.4 km while haplotype components reach 39.7 km (SparsePainter); global-PC linear models fail in African and admixed cohorts where explicit ancestry-disentangled representations succeed (DisPred — though no reported AUC value is given on Lacuna for it, only its qualitative advantage over PRS-Lasso and standard networks); and in admixed individuals the relevant ancestry is local and chromosomal, not global. Our design uses local ancestry covariates plus within-donor allele-specific expression, which is immune to stratification by construction. We still run the top-10-PC model as a covariate-set ablation, so the comparison is empirical rather than rhetorical.

Why not just fine-map to the single causal variant using the diverse cohort? This is the deepest alternative because it is the field's dominant strategy, and our transfer experiment is designed as a direct duel with it rather than a dismissal: we compare replication rates of tag-SNP, fine-mapped single-variant, and haplotype-block representations at matched discovery stringency. If the compound-haplotype view is wrong for dynamic traits, the fine-mapped variant will replicate at least as well and the proposal's third claim collapses — but the PIN1 cis-synergy result, where neither mutation alone does anything, is the concrete reason to expect haplotypes to win at some loci.

Decisive Checks

The proposal stands or falls on three numbers. First, the head-to-head discovery yield against the dynamic single-SNP baseline at matched FDR, with every claimed haplotype-only effect validated by within-donor allele-specific expression. Second, the cross-ancestry replication gradient — haplotype versus fine-mapped SNP versus tag SNP — in held-out African, East Asian, and admixed donors; a null or inverted gradient is a decisive negative result that would resolve the field's fine-mapping-versus-haplotype contradiction the other way. Third, the 3D-versus-2D ablation on projected traits, which either justifies or indicts the volumetric imaging investment. The GEUVADIS replication (22 associations, under-4% switch-error sensitivity loss) gates everything upstream. No single experiment in the file has run any of these comparisons, because no existing dataset contains multi-donor genotypes, dense time courses, and volumetric imaging simultaneously — assembling that dataset and its analysis model is the contribution.