When a person carries one broken copy of a gene, every cell in the body carries it, yet in Parkinson's disease only the dopamine-making neurons of one small midbrain nucleus die, while heart and kidney cells that use the same genes every day survive untouched. Plant biologists have spent a century on a structurally identical puzzle: cross two inbred parents and the offspring are not the average of the parents — sometimes dramatically better (hybrid vigor, or heterosis), sometimes catastrophically worse. A large part of the answer, worked out with RNA sequencing in rice and maize hybrids, is allele-specific expression: in a heterozygote, a cell does not read its two gene copies equally, and which copy it preferentially transcribes — decided largely by short non-coding regulatory switches sitting next to the gene — determines whether the organism gets vigor or metabolic collapse. Nobody has ever asked whether the same "which copy gets read" logic explains why one cell type dies and its neighbor does not in a neurodegenerative brain.
I propose the lost-masking hypothesis: resilient cell types execute an allelic masking maneuver — their local regulatory grammar skews transcription toward the intact allele of a heterozygously damaged gene, the same dominance mechanism long invoked for heterosis — while vulnerable neuronal subtypes, whose enhancer grammar differs, express both copies near-equally (or even skew toward the damaged one) precisely in the genes whose dosage they can least afford to lose. This mechanism is invisible to whole-exome sequencing, which sees only the broken code and not the regulatory decision made about it, and it is invisible to standard bulk RNA-seq, which averages across cell types and across the two chromosomes. Testing it requires the one measurement that connects all four literatures: allele-resolved single-cell RNA sequencing on personal diploid genomes.
Core Claim
Selective neuronal vulnerability is, in part, a cis-regulatory dosage phenomenon, not purely a demand phenomenon. The field's current explanation is the "canary in the coal mine" model: ubiquitously expressed housekeeping genes, when damaged, kill the most metabolically and axonally stressed neurons first (Whole-exome sequencing and burden analysis identify six novel candidate risk genes and expand the genetic landscape of Parkinson's disease; Biophysical Properties of Somatic and Axonal Voltage-Gated Sodium Channels in Midbrain Dopaminergic Neurons). That model explains why the most stressed cells go first, but not why other equally stressed non-neural cells are spared. I claim the spared cells are spared partly because they mask. Formally, for a heterozygous locus in cell type $c$, let $r_c$ be the fraction of transcripts from the intact allele; the masking score is $M_c = \log_2\!\big(r_c/(1-r_c)\big)$, and the hypothesis predicts $\Delta M = M_{\text{vulnerable}} - M_{\text{resilient}} < 0$ at risk loci, within the same donor, driven by cell-type-specific non-coding grammar rather than by the coding lesion itself. The prediction extends to non-coding risk variants, which is where the exome-centric critique bites: the Parkinson's exome study itself concedes it missed all promoters, enhancers, and untranslated regions, and found no significant association from missense variants — only protein-truncating ones (Whole-exome sequencing and burden analysis...) — so any mechanism acting through graded dosage rather than catastrophic truncation is entirely absent from current risk catalogs.
The heterosis literature supplies the conceptual scaffold, and one sobering precedent. In rice F1 hybrids, roughly 30% of genes show parental allelic imbalance, driven predominantly by local cis-regulatory variation rather than imprinting (Global RNA sequencing reveals that genotype-dependent allele-specific expression contributes to differential expression in rice F1 hybrids), while a second super-hybrid rice study reports pervasive maternal bias (Transcriptome profiling of two super hybrid rice provides insights into the genetic basis of heterosis) — an unresolved contradiction I treat as a design requirement: both cis and parent-of-origin effects must be separable in the analysis. Most importantly, the chicken negative-heterosis study shows what a failed non-additive decision looks like: F1 crosses with inferior growth show a non-additive transcriptional slump specifically in oxidative phosphorylation genes, producing mitochondrial energy failure in the most energy-demanding tissue, skeletal muscle (Genetic basis of negative heterosis for growth traits in chickens revealed by genome-wide gene expression pattern analysis). That is a precise molecular template for what I predict happens in dopaminergic neurons carrying heterozygous damage to mitochondrial-maintenance genes such as OPA1, which had the highest carrier frequency (0.17%) among Parkinson's patients in the exome study.
Research Program
The program has three strands. First, measurement: build allele-resolved single-cell transcriptomics on personal diploid genomes for matched vulnerable and resilient cell types. The methodological core is the pipeline from the uniform allele-specific expression survey, which eliminated reference-mapping bias by mapping reads to each individual's own diploid genome and modeled read overdispersion with a beta-binomial, detecting 63,541 allele-specific expression variants across 382 individuals — but only in lymphoblastoid cell lines, never in neurons (A uniform survey of allele-specific binding and expression over 1000-Genomes-Project individuals). I will port that pipeline to single-cell resolution, which matters because the existing single-cell disease work is explicitly non-allelic: the Alzheimer's atlas (over 7 million nuclei from 84 donors across 10 regions) catalogued subtype-selective loss of intratelencephalic projection neurons and VIP/SST/PVALB interneurons but modeled expression additively across both chromosomes (An integrated, multiregional, and multimodal cell atlas of Alzheimer's disease), and the spinal motor neuron study showed limb-innervating pools are molecularly more stress-sensitive than trunk pools without any allele resolution (Single-cell transcriptomic analysis reveals diversity within mammalian spinal motor neurons).
Second, prediction: a two-tier model that combines the strengths of the existing non-coding variant tools while fixing their documented blind spots. DeepVRegulome fine-tuned a DNA language model on over 700 regulatory datasets and reaches 94–96% accuracy on splice-site disruption and over 85% on transcription-factor binding changes, but only within 300-base-pair local windows (DeepVRegulome: DNABERT-based deep-learning framework for predicting the functional impact of short genomic variants on the human regulome). CREaTor links candidate regulatory elements across 2-megabase windows to target-gene expression with a hierarchical transformer, reaching a mean Pearson correlation of 0.902 in an unseen cell line (CREaTor: zero-shot cis-regulatory pattern modeling with attention mechanisms), but it predicts expression from reference epigenomes, not allele balance on personal genomes. I will extend the CREaTor architecture to take diploid haplotypes and cell-type-specific epigenomes as input and output a predicted masking score per allele — the quantity no current model produces, and the gap the single-cell foundation-model survey itself flags as the "tokenization gap," since models like Geneformer and scGPT tokenize only protein-coding gene symbols (A Survey on Foundation Language Models for Single-cell Biology).
Third, manipulation: validate predicted masking enhancers with massively parallel reporter assays and CRISPR base editing, following the design logic of the machine-guided cis-regulatory element work, whose Malinois predictor achieves Pearson r = 0.88–0.89 against empirical reporter activity and whose designed elements drove cell-type specificity in 94.1% of constructs in vitro (Machine-guided design of cell-type-targeting cis-regulatory elements). That paper designed synthetic enhancers for cell-type targeting; I use its machinery on naturally occurring heterozygous haplotypes to ask a question it never posed — whether the two haplotypes of one person differentially activate in vulnerable versus resilient neurons.
One ethical constraint shapes the data plan: allele-specific expression data is re-identifying. A list of an individual's allele-skewed genes alone matched 55% of individuals directly to their genomes in a 2,504-person database, and 80% within the top 20 matches (Recovering genomes and phenotypes using allele-specific gene expression). All allele-resolved neuronal data must therefore live in controlled-access repositories, not open ones.
Baselines
The naive reference point is the assumption every existing disease atlas makes implicitly: allelic expression is balanced, so total transcript abundance is the only quantity worth modeling. Under this null, $M_c = 0$ for every cell type, and vulnerability must be explained entirely by demand. Beating it is straightforward in form but decisive in content: significant, donor-reproducible departure of $M_c$ from zero at risk loci in vulnerable subtypes only, with the reproducibility bar set by the family-trio heritability of allelic behavior measured in the uniform survey, $r \approx 0.78$ to $0.84$ (A uniform survey of allele-specific binding and expression...) — if cell-type-specific skew is a real biological trait it should show similar within-donor consistency across differentiations.
The strongest existing competitor for risk discovery is the exome-wide protein-truncating-variant burden analysis of 3,602 Parkinson's cases and 145,496 UK Biobank controls, which identified six novel risk genes (RGMB, SNX13, MGST2, NMBR, RCBTB1, JAG1) with odds ratios from 5.6 to 16.6, and RGMB at odds ratio 12.54 (Whole-exome sequencing and burden analysis...). It is the strongest competitor because it is the only analysis with large, replicated effect sizes on exactly my loci and phenotype. The comparison will be run on the same carriers and the same phenotype labels: burden status predicts who is a case; my masking score must add within-carrier discrimination — separating carriers whose vulnerable neurons lose masking from carriers who mask well, which burden testing cannot do in principle because it collapses the two haplotypes into one indicator variable. If masking adds no within-carrier predictive layer, the proposal fails even if the biology is real.
The strongest measurement competitor is the uniform ASE survey pipeline itself, run on my same neuronal data but without cell-type stratification — pooled across all neurons the way its authors pooled lymphoblastoid cells. Same reads, same personal-genome mapping, same beta-binomial tuning, same multiple-testing correction; the only difference is stratification by subtype. Beating it means finding skew at loci where pooled analysis sees none, and the pass condition is directional: the vulnerable subtype, not the pooled sample, carries the anti-masking signal. Two ablations isolate the key ingredients: mapping to the reference genome instead of personal diploid genomes (isolates the mapping-bias correction), and single-cell data collapsed to bulk before ASE calling (isolates subtype resolution). Each should degrade the result if the corresponding ingredient matters.
Why not just…?
Why not just scale up the exome burden study? Because the coding-only strategy has already delivered what it can and its own authors name the ceiling: the association is driven exclusively by protein-truncating variants, missense variants showed no significant association, and the study explicitly excludes all non-coding drivers (Whole-exome sequencing and burden analysis...). Common-variant GWAS in the same literature carries odds ratios of only 1.1 to 1.5, leaving most heritability unexplained. More fundamentally, a burden score is cell-type-blind: it can never say which neurons die, which is the entire question selective vulnerability poses. My proposal does not compete with burden testing for gene discovery; it competes for the mechanism of cell-type specificity.
Why not just do bulk RNA-seq on patient brain tissue? Bulk sequencing demonstrably finds pathway-level signals — the chicken study located the oxidative phosphorylation slump in bulk skeletal muscle (Genetic basis of negative heterosis...) — but the heterosis field's own limitations section shows why bulk fails here: it averages out both cell-type-specific effects and the two alleles, since allele-resolved analysis in hybrids required tracking parental reads in a common nucleus (Global RNA sequencing reveals...). Post-mortem bulk brain also confounds expression change with cell loss — the Alzheimer's atlas showed subtype-specific neuronal death that bulk tissue would misread as downregulation (An integrated, multiregional...). The within-donor, within-cell-type allelic contrast I need simply cannot be constructed from bulk samples.
Why not just apply DeepVRegulome or CREaTor off the shelf? Both fall short of this specific question in documented ways: DeepVRegulome scores variants only in 300-base-pair windows, missing the long-range enhancer contacts that span up to 2 megabases, and it was validated in dividing glioblastoma cells, not non-dividing neurons (DeepVRegulome...); CREaTor predicts expression from reference epigenomes and has never modeled a diploid genome or an allelic imbalance (CREaTor...). But I am honest about the threat: if CREaTor's zero-shot machinery, transferred to neuronal epigenomes, already predicts allele balance without my diploid extension, then my contribution reduces to an application study. The held-out-locus prediction experiment below is designed to settle exactly this — my model must beat both off-the-shelf tools at predicting the direction of allelic skew, not just expression level, and the ablation that removes the diploid input isolates what is actually new.
Why not just run a standard eQTL study in GTEx? GTEx is the natural resource for linking variants to expression, but it is bulk post-mortem tissue; no reported result found on Lacuna for allele-specific expression measured at the resolution of substantia nigra dopaminergic neurons in GTEx, and the best existing human ASE catalog comes from lymphoblastoid cell lines (A uniform survey...) — a cell type with no relevance to neuronal enhancer grammar. An eQTL would also conflate the two haplotypes' joint effect, whereas the hypothesis is specifically about their imbalance within one cell type.
Proposed Experiments and Decisive Checks
The discovery experiment uses isogenic induced pluripotent stem cell lines: CRISPR heterozygous knock-ins of protein-truncating variants in OPA1, SNX13, and RGMB in a healthy background, differentiated in parallel into midbrain dopaminergic neurons (the vulnerable type) and cortical glutamatergic neurons and astrocytes (resilient comparisons), following the matched multi-lineage design of the 22q11.2 deletion study, which combined morphological phenotyping with transcriptomics across neurons, progenitors, and astrocytes from the same genetic background (Combining phenomics with transcriptomics reveals cell-type-specific morphological and molecular signatures of the 22q11.2 deletion). Each line gets whole-genome sequencing to build its personal diploid reference, then allele-resolved single-cell RNA-seq with the beta-binomial pipeline, yielding $M_c$ per expressed heterozygous locus per cell type. The primary endpoint is $\Delta M$ at the edited loci: dopaminergic neurons should show significantly lower masking than the sibling cell types from the same clone. A secondary endpoint benchmarks scale: heterosis studies find roughly 30% of genes allelically biased in plant hybrids (Global RNA sequencing reveals...); I will report the genome-wide biased fraction in each human cell type as the first such number for neurons — no reported result found on Lacuna for genome-wide allele-specific expression fractions in any human neuronal subtype.
A stress experiment tests the dynamics: dopaminergic neurons operate with an exceptionally narrow biophysical reserve (Biophysical Properties...), and the chicken precedent shows the collapse arrives under load (Genetic basis of negative heterosis...). I will measure $M_c$ before and after mild mitochondrial stress in both cell types; the hypothesis predicts stress erodes masking in dopaminergic neurons specifically, converging on the oxidative phosphorylation gene set.
The prediction experiment evaluates the two-tier diploid model on loci and cell types held out from training, reporting the correlation between predicted and observed allelic ratios and the area under the receiver operating characteristic for skew direction. The reference points are explicit: CREaTor's 0.902 Pearson correlation for expression prediction in an unseen cell line and its roughly 0.84 area under the curve against the Activity-by-Contact baseline for enhancer–gene pairing (CREaTor...), plus DeepVRegulome's local-only scores (DeepVRegulome...). The diploid model must beat both at skew direction; matching them on expression level while failing on skew would falsify the modeling contribution even if the biology holds.
The manipulation experiment closes the loop: massively parallel reporter assays comparing the two haplotypes of the masking enhancers identified above, in dopaminergic neurons versus astrocytes, with Malinois used as the pre-screen predictor at its reported r = 0.88–0.89 fidelity (Machine-guided design...); then CRISPR base editing of the causal regulatory variant in the iPSC line. The decisive check is whether editing restores masking in dopaminergic neurons and, with it, resistance to mitochondrial stress — the direct analog of showing that fixing the regulatory decision, not the coding lesion, rescues the vulnerable cell.
Two limitations are acknowledged rather than argued away. Nuclear single-cell methods miss axonal biology, where Huntington's aggregates and transport blockade begin long before somatic death (Huntingtin Aggregate-Associated Axonal Degeneration is an Early Pathological Event in Huntington's Disease Mice); my design partially evades this by using intact iPSC-derived cells rather than isolated nuclei, but distal neurite translation remains out of reach and is left to future work. And the rice-versus-super-hybrid contradiction over cis-driven versus parent-of-origin skew (Global RNA sequencing reveals... versus Transcriptome profiling of two super hybrid rice...) means imprinting effects must be explicitly tested and reported, not assumed away. If the isogenic carrier lines show balanced allelic expression in dopaminergic neurons, indistinguishable from resilient siblings and from unedited controls, the lost-masking hypothesis is wrong for these loci — and the field should return its full attention to the biophysical demand model, now with one candidate mechanism excluded rather than merely untested.
References
- Whole-exome sequencing and burden analysis identify six novel candidate risk genes and expand the genetic landscape of Parkinson’s disease
- Biophysical Properties of Somatic and Axonal Voltage-Gated Sodium Channels in Midbrain Dopaminergic Neurons
- Global RNA sequencing reveals that genotype-dependent allele-specific expression contributes to differential expression in rice F1 hybrids
- Transcriptome profiling of two super hybrid rice provides insights into the genetic basis of heterosis
- Genetic basis of negative heterosis for growth traits in chickens revealed by genome-wide gene expression pattern analysis
- A uniform survey of allele-specific binding and expression over 1000-Genomes-Project individuals
- An integrated, multiregional, and multimodal cell atlas of Alzheimer’s disease
- Single-cell transcriptomic analysis reveals diversity within mammalian spinal motor neurons
- DeepVRegulome: DNABERT-based deep-learning framework for predicting the functional impact of short genomic variants on the human regulome
- CREaTor: zero-shot cis-regulatory pattern modeling with attention mechanisms
- A Survey on Foundation Language Models for Single-cell Biology
- Machine-guided design of cell-type-targeting cis-regulatory elements
- Recovering genomes and phenotypes using allele-specific gene expression
- Combining phenomics with transcriptomics reveals cell-type-specific morphological and molecular signatures of the 22q11.2 deletion
- Huntingtin Aggregate-Associated Axonal Degeneration is an Early Pathological Event in Huntington's Disease Mice