A common Alzheimer's risk variant near the tau gene, rs2471738 in MAPT, is linked to higher MAPT expression in human hippocampus. That link comes from bulk tissue: ground-up brain in which neurons, glia, and blood vessels are averaged together. In Alzheimer's disease the neurons that matter die early while glia expand. A bulk association can therefore mean either of two things. The variant may change how much tau RNA vulnerable neurons make, or it may only track how many neurons survive in a sample. We propose to separate these explanations with a measurement that cell composition cannot fake: the ratio of the two alleles' tau pre-mRNA inside one donor's nuclei, sorted by cell type. A heterozygous nucleus carries both alleles in the same cell, so any imbalance between them reflects regulation acting on that copy of the gene, not a shift in which cells are present.
The second half of the problem is design. Post-mortem samples are scarce, donors differ sharply in disease stage, and quantitative PCR — the lab assay that counts transcripts by tracking fluorescence over amplification cycles — introduces plate-to-plate drift. We treat the assignment of donors to PCR plates as a combinatorial covering problem. Each plate holds a closed chain of donors in which every neighbouring pair differs in genotype and is matched on tau pathology. Covering all donors with such chains is exactly the cycle-factor problem from extremal graph theory, which reduces to perfect matching in a hypergraph.
Core Claim
The core claim is that a within-donor, cell-type-sorted allelic ratio, combined with a covariate-matched across-donor design, can tell whether a bulk neurodegeneration eQTL acts inside the cells that degenerate. Here an eQTL (expression quantitative trait locus) is a DNA variant whose genotype predicts a gene's RNA level. The ratio removes composition confounding for the question of whether the variant acts on its own gene copy. The matched design is needed for what the ratio cannot answer: the effect in homozygous carriers, and whether the effect changes with tau burden. A genotype-by-pathology interaction is the quantity that matters for selective vulnerability.
The motivating evidence is A comprehensive analysis of MAPT-related genetic risk in Alzheimer's disease. It meta-analysed 28,618 people, reported an allelic OR of 1.15 for rs2471738, and used GTEx bulk hippocampus to show that CT carriers express more MAPT than CC carriers at P = 0.04. The authors explicitly call for functional validation. The atlas in An integrated, multiregional, and multimodal cell atlas of Alzheimer's disease supplies the reason to doubt the bulk reading. Across 84 donors and more than 7 million nuclei, limbic regions accumulate phosphorylated tau early, and L2/3 and L5 intratelencephalic excitatory neurons and SST, PVALB and VIP interneurons are lost, alongside glial reactivity. That is precisely the composition shift that can produce a spurious bulk eQTL. The weak bulk P-value is itself a reason to run this study: a clear null in sorted neurons would be an informative outcome.
What We Rejected First
The obvious idea at this intersection is to infer gene regulatory networks in vulnerable neurons and extract their directed feedback loops as a cycle-factor. A cycle-factor is a set of vertex-disjoint directed cycles that together cover every vertex. We dropped this idea. Genes sit in many overlapping loops, so forcing a partition into disjoint cycles imposes a constraint biology does not obey. A reviewer could also fairly say that simple motif counting with enrichment tests answers the same question. Neither Gene Regulatory Network Inference as Relaxed Graph Matching, which recasts network inference as continuous matching and ignores edge direction, nor Inferring gene regulatory networks by hypergraph generative model, which treats cells as hyperedges in a variational autoencoder, would be improved by such a partition.
The cycle-factor reduction does fit somewhere else: allocating samples. There, disjointness is a real physical constraint, because each scarce nuclei sample goes onto exactly one plate.
Research Program
The first stage is computational and runs on the atlas. If donor genotypes are not released, we call rs2471738 from reads covering the site. The SNP is intronic, and single-nucleus RNA carries abundant unspliced pre-mRNA, so the site should be covered in snRNA-seq and in chromatin-accessibility reads. Per cell type, we fit the zero-inflated negative binomial model of SCeQTL: an R package for identifying eQTL from single-cell parallel sequencing data. It separates two genotype effects: one on whether MAPT is detected in a nucleus at all, and one on its level when it is detected. We fit this alongside the per-cell-type pseudobulk linear models standard in Robust Linear Models for Cis-eQTL Analysis. In heterozygous donors we then count reference and alternate reads at the SNP within each cell type. This gives a single-nucleus estimate of allelic imbalance in pre-mRNA, which reflects transcription rate directly.
The second stage is experimental and uses an independent genotyped hippocampal cohort from a brain bank. Nuclei are sorted into neuronal and major glial fractions. Each fraction is assayed by RT-qPCR for total MAPT, intronic MAPT pre-mRNA, and an allele-discriminating probe pair at rs2471738. The probe pair reports the alternate-allele fraction in heterozygotes, and that fraction needs no reference gene. For across-donor comparisons we run a 96-gene panel. This clears the roughly 50-gene minimum that Data-driven normalization strategies for high-throughput quantitative RT-PCR identifies for quantile normalization. That paper found quantile normalization reached a 3.36% average coefficient of variation, against 3.60% for GAPDH and 4.77% for raw data, in THP-1 cells. Its assumptions may fail in degenerating neurons, where global transcription shifts. We therefore report all three normalizations next to the allele ratio, which is immune to that failure.
The third stage is the allocation design. We build a directed compatibility graph $D$ on samples. An arc $u \to v$ is present when $v$'s genotype follows $u$'s in the cyclic order CC→CT→TT→CC, and when $u$ and $v$ agree within tolerance on Braak-like tau stage, age, and neuronal fraction. A directed 6-cycle in $D$ is then a plate block holding two donors of each genotype, in which every adjacent pair is matched. This chain condition is weaker than requiring all pairs to match, and so is feasible far more often. Following Cycle-factors in oriented graphs, covering the samples by such cycles is a perfect matching in the 6-uniform hypergraph $H(D)$, whose hyperedges are the directed 6-cycles. We solve
$$\max \sum_{C} x_C \quad \text{s.t.} \quad \sum_{C \ni v} x_C \le 1 \ \ \forall v, \qquad B x \le d, \qquad x_C \in \{0,1\},$$
where $x_C$ indicates that cycle $C$ becomes a plate, the first constraint keeps each sample on one plate, and the few rows $B x \le d$ bound global sex and region balance. Those extra rows are a constraint backdoor in the sense of Parameterized Algorithms for Matching Integer Programs with Additional Rows and Columns, which shows that such rows make the problem W[1]-hard in general. This result constrains the design rather than supplying the method. At cohort sizes below about 150 we use a generic integer-programming solver and report solve times. Uncovered samples are reported, not dropped silently. The large-$n$ semi-degree thresholds of the cycle-factor theorems are not usable guarantees at this scale. We use them only as a diagnostic, checking how compatibility density relates to feasibility in practice.
Baselines
For the scientific question, the naive reference point is the published bulk GTEx hippocampal association at P = 0.04 (A comprehensive analysis of MAPT-related genetic risk in Alzheimer's disease), reproduced as bulk RT-qPCR with GAPDH normalization on unsorted tissue from our cohort. The composition-adjusted variant of this baseline adds the estimated neuronal fraction as a covariate. It is the cheap fix a reviewer would expect, and it is the comparison that most directly tests whether sorting adds anything.
The strongest existing single-cell method is SCeQTL (SCeQTL: an R package for identifying eQTL from single-cell parallel sequencing data). Its authors report AUC higher than or comparable to Matrix eQTL on simulations, but we found no reported numerical result on Lacuna for SCeQTL. We therefore re-run it ourselves, on the same atlas nuclei, cell-type labels, and covariates as the pseudobulk linear model and our allelic-ratio test. All three methods share one permutation scheme for calibration. A method is considered better if it gives tighter confidence intervals on the per-cell-type effect at matched false-positive rate under donor-label permutation. It must also agree better with the independent RT-qPCR allele ratios.
For the design question, the baselines are complete randomization of donors to plates, stratified randomization blocked on genotype only, and greedy pairwise nearest-neighbour matching. Each baseline receives the same plate count and capacity. The ablation that isolates our key ingredient keeps the cycle-factor structure but removes the pathology tolerance from arcs. This leaves cycles balanced on genotype alone, so any gain can be attributed to within-chain covariate matching rather than to the cycle structure. The cycle design beats the baselines if it reduces root-mean-square error of the genotype-by-tau-stage interaction estimate in the simulation below. We set the bar in advance at a reduction of at least 20% against stratified randomization.
Why not just…?
Why not just block plates by genotype and randomize the rest? This is the real threat to the design component. In a large cohort with weak pathology effects, blocking would probably suffice. The case for matching rests on survivorship. Vulnerable neurons are depleted in late-stage donors, so tau stage confounds both composition and expression. With roughly a hundred donors and a rare TT class, chance imbalance in stage within plates is likely. The simulation experiment decides this question directly. If stratified randomization comes within the 20% bar, we drop the cycle design and say so.
Why not just run SCeQTL or pseudobulk regression on the atlas and skip the wet lab? Single-nucleus data drop many transcripts, and the atlas has only 84 donors, so an across-donor single-cell eQTL may be underpowered at a variant whose bulk signal was only P = 0.04. Across-donor single-cell regressions also still compare different people, who differ in disease stage. The allele ratio in sorted nuclei is a within-person comparison. RT-qPCR measures it at higher depth on an independent cohort, which also guards against the atlas result being an artefact of read calling at a single site.
Why not just measure allele-specific expression and skip the across-donor design? If the only question were whether the variant acts on its own gene copy in neurons, the allelic ratio would be enough, and we treat it as the primary test. It cannot estimate the homozygous effect, and it cannot test whether the allelic effect grows or shrinks with tau burden without comparing donors at different stages. That interaction is the vulnerability-relevant quantity, and the matched design exists to estimate it.
Why not just use an off-the-shelf matching package or a solver directly? We do use a generic solver. The contribution of the hypergraph reduction is the formulation: chain-matched rather than clique-matched blocks, written as an exact covering problem with a few balance rows. Cyclic orientation by genotype forces each block to be balanced without enumerating cliques. We will report how often clique-matched blocks are infeasible at the same tolerances, which shows whether the chain relaxation is necessary.
Proposed Experiments
The first decisive check is a simulation built from real atlas donors. We resample donors with their actual tau stages and cell-type proportions, plant a known allelic effect in excitatory neurons, optionally scaled by tau stage, and generate bulk, sorted, and allele-ratio measurements with plate effects drawn at the measured coefficient-of-variation scale. Each design and estimator is scored on bias, root-mean-square error, and coverage for the main effect and the interaction. A null scenario with no cell-intrinsic effect but realistic neuronal loss measures how often bulk analysis produces a false eQTL. If composition-adjusted bulk analysis recovers the planted effect as well as the allele ratio does, the sorting step is unnecessary, and we will report that.
The second check applies SCeQTL, pseudobulk regression, and the allelic-ratio test to rs2471738 and MAPT in each atlas cell type, in hippocampus and entorhinal cortex, with a neocortical region as an amyloid-first contrast. The prediction from the bulk result is a T-allele excess in excitatory neuron nuclei. A shift visible only between donors and absent within donors would point to composition.
The third check is the sorted-nuclei RT-qPCR cohort. We measure allele fractions in heterozygotes per fraction, and across-donor effects under the cycle-factor design versus stratified randomization on split halves of the cohort. The concordance of the RT-qPCR allele fractions with atlas read-based fractions, cell type by cell type, is the final test of whether the hippocampal MAPT eQTL belongs to the neurons that die.
References
- A comprehensive analysis of MAPT-related genetic risk in Alzheimer’s disease
- An integrated, multiregional, and multimodal cell atlas of Alzheimer’s disease
- Gene Regulatory Network Inference as Relaxed Graph Matching
- Inferring gene regulatory networks by hypergraph generative model
- SCeQTL: an R package for identifying eQTL from single-cell parallel sequencing data
- Robust Linear Models for Cis-eQTL Analysis
- Data-driven normalization strategies for high-throughput quantitative RT-PCR
- Cycle-factors in oriented graphs
- Parameterized Algorithms for Matching Integer Programs with Additional Rows and Columns