Protein aggregation diseases like Alzheimer's and Parkinson's are governed by a two-step physical mechanism: proteins first collapse into dense, disordered, liquid-like clusters, and only then — slowly, inside those clusters — do they reorganize into the ordered fibrils that damage tissue. Separately, genetic epidemiology tries to infer whether such molecular processes causally drive disease using Mendelian randomization, which treats natural genetic variants as randomized proxies for a modifiable exposure. That machinery, however, was built almost entirely on European-ancestry data, and its instruments — genetic variants chosen because they tag an exposure in one population — routinely fail in other populations because the correlation structure between tag variants and functional variants differs across ancestries.
This proposal argues these two failures are connected and jointly fixable. The molecular phenotypes that matter for two-step nucleation are not monolithic "disease risk" but fine-grained quantities such as the ratio of splice isoforms of a protein — and the physics is sharply non-linear, because aggregation switches on only when concentrations and isoform mixtures cross thresholds. Both properties interact badly with standard practice: European-discovered tag variants for a splice-regulating variant may be uncorrelated with it in African or East Asian populations, and the linear dose–response assumed by standard two-sample Mendelian randomization misrepresents a threshold process. We propose to build stage-separated, cross-ancestry genetic instruments for the splice-isoform mediators of two-step nucleation, calibrate the non-linear isoform-to-kinetics map with in vitro data, and test causal transfer across four biobanks.
Core Claim
The core claim is twofold. First, ancestry-biased causal inference for aggregation diseases fails partly because genetic instruments are blind to the stage structure of nucleation: variants that shift protein isoform ratios act on the condensation and seeding stage, not on the later ordering stage, and European tag variants for these splicing effects lose relevance in other ancestries. Second, a stage-separated design — cross-ancestry fine-mapped splicing instruments as the first stage, an in vitro-calibrated kinetic map as the link function, and disease risk as the second stage — yields causal estimates that transfer across ancestries where monolithic, European-derived instruments attenuate to false negatives.
The anchor system is α-synuclein, the Parkinson's disease protein encoded by SNCA. Aβ42 promotes the aggregation of α-synuclein splice isoforms via heterogeneous nucleation proved in vitro that Aβ42 fibrils seed α-synuclein aggregation through a strict two-step cascade, that seeding is unidirectional, and that the kinetics depend exquisitely on splicing: isoforms lacking exon 5 (αSyn-112, αSyn-98) aggregate much faster, with pre-formed Aβ42 seeds cutting their half-times by roughly 40%, and isoform mixtures as dilute as 1–10% αSyn-112 accelerating full-length αSyn-140 aggregation. Meanwhile, Cross-ancestry atlas of gene, isoform, and splicing regulation in the developing human brain showed that isoform and splicing regulation explains more disease heritability than bulk gene expression, and that diverse ancestries refine causal variant sets dramatically (at MTFR1, from ~50 candidates to 3–5). Nobody has connected the two: whether ancestry-specific splicing regulation changes the baseline ratio of exon-5-lacking α-synuclein isoforms — and hence the genetic driving force for Aβ42-catalyzed heterogeneous nucleation — is entirely unmeasured.
Research Program
The program has three workstreams.
Workstream 1: cross-ancestry fine-mapping of stage-specific instruments. Using the multi-ancestry regulatory atlas (672 fetal brain samples; 45% European, 25% Hispanic/Latino, 22% African, 8% East/Southeast Asian) together with UK Biobank and BioBank Japan splicing and isoform summary statistics, we will run XMAP — a cross-population fine-mapping tool that exploits the shorter linkage-disequilibrium blocks of African genomes to triangulate causal variants — on the SNCA locus, separating splicing QTLs that modulate the exon-5 isoform ratio from expression QTLs that shift total protein abundance. XMAP is the methodological supplier here: it identified 149 putative causal variants for LDL cholesterol where single-population SuSiE found 52, and it absorbs hidden stratification through LD-score-regression intercepts, which matters because stratification is the classic confounder of cross-population analysis. As a secondary system we will fine-map FUS (ALS), where Nanoscale profiling of evolving intermolecular interactions in ageing FUS condensates showed that the two nucleation steps are physically separable — a surface disorder-to-order transition with a β-sheet infrared signature at 1625 cm⁻¹ appearing at 4–8 hours, distinct from core gelation driven by intrinsically-disordered-region contacts, with stiffness rising from ~10 kPa to ~1 MPa over 24 hours. This lets us ask whether human variants cluster preferentially in intrinsically disordered regions (condensation stage) versus structured cores (ordering stage), a question no fine-mapping study has addressed.
Workstream 2: instrument transfer diagnostics. For every instrument discovered in Europeans, we will compute the correlation ($r^2$) between the tag variant and the fine-mapped causal splice variant in African and East Asian reference panels, and measure how this attenuation predicts the degradation of Mendelian-randomization estimates in non-European outcomes. This mechanizes the failure mode documented by Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing, where thresholds built in White cohorts simply failed elsewhere, and by Discovery and fine-mapping of adiposity loci using high density imputation of genome-wide association studies in individuals of African ancestry, where African-ancestry recombination uncoupled false tag associations from true functional drivers.
Workstream 3: a calibrated two-step estimator. Standard two-sample Mendelian randomization assumes a linear chain from instrument to exposure to outcome. Two-step nucleation is instead a threshold process, and Unveiling Challenges in Mendelian randomization for Gene-Environment Interaction proves that off-the-shelf two-stage estimators (predictor substitution and residual inclusion) are both biased for binary outcomes through non-collapsibility of odds ratios. Our design sidesteps this by keeping the first stage linear (genotype → isoform fraction $f$, a continuous quantity) and pushing the non-linearity into an explicit, physically calibrated link function $g(f)$, the aggregation propensity. The ancestry-specific estimand for a splice instrument $Z$ is then
$$\beta^{(a)}_{Z \to Y} \;=\; \beta^{(a)}_{Z \to f}\;\cdot\; g'(f_a)\;\cdot\; \lambda,$$
where $\beta^{(a)}_{Z \to f}$ is the ancestry-specific effect of the instrument on the exon-5-lacking isoform fraction, $g'(f_a)$ is the slope of the calibrated kinetic map at that ancestry's mean isoform fraction $f_a$, and $\lambda$ converts nucleation propensity to disease-risk scale. The map $g$ is fitted from in vitro kinetics: we will extend the AmyloFit 2.0 experiments of Aβ42 promotes the aggregation of α-synuclein splice isoforms via heterogeneous nucleation — which already resolved half-times, seeding thresholds (active at 0.1 µM Aβ42), and isoform-mixture effects — to a denser grid of isoform ratios spanning the population-plausible range, measuring aggregation half-time $t_{1/2}(f)$ by ThT fluorescence and fitting $g$ to the resulting curve. This is the proposal's key ingredient: the non-linearity is not a nuisance parameter to be robustified away but a physically measurable function.
Baselines
The do-nothing reference point is the prevailing practice: generalized summary-data-based Mendelian randomization on monolithic clinical endpoints in European data, as in Causal associations between risk factors and common diseases inferred from GWAS summary data, which resolved 45 causal associations in up to 405,072 individuals (e.g., one standard deviation of BMI conferring a 3.29-fold increase in type-2 diabetes risk) but, by its authors' own statement, used exclusively European cohorts. Beating this baseline does not mean finding more associations in Europeans; it means producing causal estimates for the SNCA isoform pathway that remain significant and directionally consistent in East Asian and African-ancestry outcomes, where the baseline approach produces no transferable estimate at all.
The strongest existing competitor is the best current cross-ancestry practice, exemplified methodologically by Mendelian randomization study shows no causal relationship between psychiatric disorders and glaucoma in European and East Asian populations: standard two-sample Mendelian randomization with European-derived instruments applied to European (UK Biobank, FinnGen, PGC) and East Asian (China Kadoorie Biobank, BioBank Japan) outcomes, with the full sensitivity battery of inverse-variance weighting, MR-Egger, weighted median, MR-PRESSO, and Steiger filtering. That study is the strongest competitor because it is the only template in the corpus that actually runs the same causal question through independent biobanks across ancestries — it found null effects for psychiatric disorders on glaucoma in both groups (European depression odds ratio 1.15, p = 0.20; East Asian 1.38, p = 0.30). We will run it on our question with the same summary statistics, the same instrument-selection p-value threshold, the same number of instruments, and the same sensitivity battery, so any difference is attributable to the stage-separated design rather than to tuning or data. We beat it if, where it reports attenuation or false negatives for splicing-mediated effects in non-European cohorts, our calibrated design recovers estimates whose signs and calibrated magnitudes agree across ancestries within confidence intervals.
Two ablations isolate the ingredients. Replacing XMAP multi-ancestry fine-mapping with single-population SuSiE isolates the LD-triangulation ingredient (XMAP found 149 versus 52 LDL variants on identical data); we expect the SuSiE-based instruments to lose cross-ancestry relevance, mirroring the MTFR1 credible-set inflation from 3–5 to ~50 variants when ancestry diversity is discarded. Replacing the calibrated link $g$ with a linear isoform-to-risk slope isolates the phase-aware ingredient; we expect the linear version to produce ancestry-dependent effect estimates that disagree in magnitude even when the underlying biology is shared, because it ignores the threshold in $t_{1/2}(f)$.
Why not just…?
Why not just run Mendelian randomization on the clinical endpoint with even more European samples? Because monolithic endpoints demonstrably mix up stages and mislead. In Causal associations between risk factors and common diseases inferred from GWAS summary data, LDL cholesterol appeared causally protective of or risk-increasing for Alzheimer's disease depending on whether 16 pleiotropic APOE-region variants were purged — the association was an artifact of treating a heterogeneous endpoint as one thing. And Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing shows that even a perfectly powered European model fails outright in other ethnic groups, so more European samples cannot buy generalizability. Our first experiment measures exactly this: the correlation of European tag variants with the causal splice variant in African and East Asian panels, which more European data leaves unchanged.
Why not just use off-the-shelf two-step Mendelian randomization for mediation, as existing two-step studies do? Papers like The causal mediating effect of smoking on the relationship between irritability and bipolar disorder chain two standard Mendelian-randomization steps through a clinical mediator, and if our mediator were linear in the outcome this would be nearly equivalent to our design — that is a real threat, not a strawman. It falls short here for two reasons. Those studies are single-ancestry, so they inherit the transfer problem rather than addressing it. More importantly, the mediator-to-outcome stage in our system is a measured non-linear threshold function, and Unveiling Challenges in Mendelian randomization for Gene-Environment Interaction proves that two-stage instrumental estimators carry persistent non-collapsibility bias for binary outcomes — their own colorectal-cancer analysis (44,500 cases, 52,235 controls) found the observational log-odds estimate of 0.14 per 5 BMI units versus 0.23 by instrumental estimation, a gap of the same order as the effects we chase. The decisive check is within our own design: if the linear-link ablation and the calibrated estimator agree within confidence intervals across all ancestries, then the phase-aware ingredient adds nothing and the simpler mediation design would have sufficed; we would report that honestly.
Why not just measure isoform ratios in the lab and skip genetics entirely? The in vitro work is necessary but not sufficient: Aβ42 promotes the aggregation of α-synuclein splice isoforms via heterogeneous nucleation is explicit that its clean-buffer system cannot say anything about isoform ratios in living humans, and the two-step mechanism's own open problem is the translational gap to crowded, chaperone-rich cellular interiors. Conversely, the genetic side cannot measure kinetics at all. Only the combination answers the population-level question — whether natural human splicing variation, distributed differently across ancestries, shifts individuals along the calibrated kinetic curve — and only cross-ancestry instruments can test whether the resulting causal claim is a human regularity rather than a European-cohort artifact.
Proposed Experiments and Decisive Checks
Experiment 1 fine-maps the SNCA locus with XMAP across the fetal-brain atlas plus UK Biobank and BioBank Japan splicing summary statistics, reporting credible-set sizes for exon-5 splicing signals versus single-ancestry SuSiE (expectation anchored by XMAP's 149-versus-52 result and the MTFR1 50-to-3–5 compression) and testing whether the causal splice variant is shared across ancestries or ancestry-specific. Experiment 2 computes the LD correlation of every European-discovered instrument with the fine-mapped causal variant in African and East Asian panels, and regresses observed cross-ancestry instrument strength against that correlation — the direct test of the relevance-transfer mechanism. Experiment 3 runs the calibrated two-step estimator on Parkinson's disease (PGC and BioBank Japan) and Alzheimer's disease (UK Biobank, FinnGen, BioBank Japan) outcomes, with the standard two-sample sensitivity battery as in the glaucoma study, comparing against the strongest baseline and both ablations under matched data and instrument budgets. Experiment 4 extends the AmyloFit 2.0 kinetic platform to a grid of isoform mixtures bracketing population-plausible ratios, fitting $t_{1/2}(f)$ and hence $g$; this extends the existing 1–10% αSyn-112 mixture data from the Aβ42/α-synuclein study rather than replacing it.
The decisive check is a two-way test. If our stage-separated, calibrated estimates agree in sign and magnitude across European and East Asian biobanks while European-derived monolithic instruments attenuate in the East Asian cohorts, the two-step mechanism is a transferable human causal pathway and the ancestry gap was an instrument-resolution problem. If the calibrated estimates still disagree across ancestries after instrument correction, then something genuinely differs in the biology — an ancestry-specific gene–environment interaction of the kind Unveiling Challenges in Mendelian randomization for Gene-Environment Interaction proves current two-stage estimators cannot resolve — and we would report that as the more important negative result. Either outcome moves past the current state, in which biophysics and population genetics study the same proteins without ever exchanging instruments.
A known limitation we will carry openly: the multi-ancestry regulatory atlas is fetal and developmental, and adult-brain splicing regulation may differ; we will treat the atlas as the discovery set and validate instrument relevance in adult tissue through isoform-level colocalization, flagging any locus where fetal and adult splicing signals disagree. No reported result found on Lacuna for adult multi-ancestry SNCA splicing effect sizes, which is itself part of the gap this proposal fills.
References
- Aβ42 promotes the aggregation of α‐synuclein splice isoforms via heterogeneous nucleation
- Cross-ancestry atlas of gene, isoform, and splicing regulation in the developing human brain
- XMAP: Cross-population fine-mapping by leveraging genetic diversity and accounting for confounding bias
- Nanoscale profiling of evolving intermolecular interactions in ageing FUS condensates
- Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing
- Discovery and fine-mapping of adiposity loci using high density imputation of genome-wide association studies in individuals of African ancestry: African Ancestry Anthropometry Genetics Consortium
- Unveiling challenges in Mendelian randomization for gene–environment interaction
- Causal associations between risk factors and common diseases inferred from GWAS summary data
- Mendelian randomization study shows no causal relationship between psychiatric disorders and glaucoma in European and East Asian populations
- The causal mediating effect of smoking on the relationship between irritability and bipolar disorder: A two-step Mendelian randomization study