Most genetic variants used to test causal claims in human health were picked from studies of people of European descent. Researchers then reuse them in African American, Hispanic/Latino, or East Asian cohorts. This proposal treats that reuse as a signal-cancellation problem borrowed from wireless engineering. A transferred variant's association is mostly "borrowed" from nearby causal variants it happens to be inherited with. That borrowed part can be partly subtracted using the target population's correlation structure. Whether pooling across ancestries helps or hurts then depends on how completely the subtraction works.
Full-duplex radios face the same tradeoff. They transmit and receive on one frequency at once, so they must subtract their own transmitted signal from what they hear. If the subtraction falls short of a sharp threshold, they do worse than simply taking turns. When Full Duplex Wireless Meets Non-Orthogonal Multiple Access puts that cliff near −110 dB of cancellation. We propose to derive, calibrate, and test the analogous cliff for cross-ancestry Mendelian randomization. Mendelian randomization is a method that uses inherited variants as natural experiments to estimate whether an exposure causes a disease. The calibration uses functionally validated causal variants as the genomic equivalent of the known pilot signals a radio uses to measure its own leakage. The deliverable is a per-locus, per-population decision rule. It says when a European-derived instrument, after cancellation, should be pooled with target-ancestry data, and when the analyst should fall back to target-only analysis.
Core Claim
Write the true causal effect estimated through instrument $t$ in target ancestry $A$. Suppose its transfer bias from linkage disequilibrium (the correlation between nearby variants inherited together) is $B$. Suppose cancellation, meaning conditioning on the target-ancestry correlation matrix weighted by a prior over which nearby variant is causal, removes a fraction $\rho$ of $B$. Then the two strategies have mean squared errors
$$\text{MSE}_{\text{pool}} = (1-\rho)^2 B^2 + \sigma_P^2, \qquad \text{MSE}_{\text{target}} = \sigma_T^2 + W^2,$$
Here $\sigma_P^2$ and $\sigma_T^2$ are the sampling variances of the pooled and target-only estimates. $W$ is the weak-instrument bias of target-only analysis, which is usually large because non-European cohorts are small. Pooling wins only when
$$\rho > \rho^* = 1 - \frac{\sqrt{\sigma_T^2 + W^2 - \sigma_P^2}}{|B|}.$$
This is the genomic counterpart of the full-duplex/half-duplex threshold. It is a "cliff" because MSE is quadratic in the residual $(1-\rho)B$. When $|B|$ is large relative to the variance gap, small shortfalls in $\rho$ flip the decision.
The new element is not the bias decomposition, which is implicit in multi-ancestry fine-mapping. The new element is measuring $\rho$ empirically. Functionally validated causal variants serve as pilots: at those loci the true causal position is known, so the realized residual is observable. We then learn a predictor of $\rho$ for loci without functional data.
Relation to Prior Work
The file reports no study joining all three directions, and no application of self-interference cancellation to genomic data. We therefore borrow the structure of the wireless work, not its physics.
Three wireless papers supply the decision structure. The FD-NOMA paper above and Duong et al. show that gains from simultaneous operation vanish below a critical cancellation depth, so simpler turn-taking wins. Improving Physical Layer Secrecy Using Full-Duplex Jamming Receivers shows that partial suppression ($\rho = 0.5$) can still pay off when the residual is concentrated in a known direction. That result motivates our second hypothesis: transfer bias may be dominated by a single leading direction of LD difference between populations, which would make partial cancellation unusually effective.
Four genetics papers motivate and constrain the work. GSMR estimates causal effects from GWAS summary statistics, which are per-variant association results published without individual genotypes. It removes pleiotropic instruments, meaning variants that affect the outcome through other routes, using the HEIDI-outlier test. Its 45 causal links come only from European cohorts. Its handling of bias is binary: an instrument is kept or deleted. We replace that with graded cancellation plus an explicit pooling decision.
HAUDI, building on GAUDI, models ancestry-specific effects using local ancestry, an estimate of which ancestral population each stretch of a person's chromosome comes from. It recovered the Duffy-null signal for white blood cell count. It is a prediction tool, though, and it needs individual-level genotypes. Our rule works on summary statistics plus a reference panel, and it targets causal effect estimation, not prediction.
Two papers supply methods we reuse. CRISPRpath runs pooled CRISPR screens that perturb many regulatory elements in parallel, giving functional ground truth. Metadata-guided Feature Disentanglement (MFD) provides sequence-based variant scores whose technical artifacts have been removed. We use them as the causal-position prior in cancellation. One constraint applies: MFD's reported FANTOM5 enhancer AUROC is only 0.68, so the prior is weak. The functional pilots are needed precisely because model scores alone will not pin down $\rho$.
Research Program
The work has three parts.
First, we build the canceller. For each European-selected instrument, we take target-ancestry LD from a reference panel and a causal-position prior. The prior is either uniform, MFD-weighted, or fixed at a known functional variant. We then estimate the instrument's expected association in the target population and subtract the borrowed component. This is a summary-statistic operation, so it fills the gap the file identifies, where HAUDI requires individual genotypes.
Second, we calibrate on pilot loci. These are loci where the causal variant has independent experimental support from CRISPR perturbation, enhancer databases (FANTOM5/Vista), or a long-established causal variant such as rs2814778. At each pilot we compute the realized $\rho$. We regress it on cheap locus features: target-versus-European LD-mismatch magnitude, the share of mismatch in its leading eigenvector (the rank-1 hypothesis), allele-frequency ratio, and MFD prior entropy.
Third, we derive and validate the threshold rule. We apply $\rho^*$ per locus and aggregate to an exposure–outcome pair. We then test whether observed estimator error changes sharply near predicted $\rho^*$, as the theory says, or degrades smoothly.
Baselines
The naive reference point is the "half-duplex" analysis: Mendelian randomization using only target-ancestry GWAS instruments. It has no transfer bias but suffers weak-instrument bias. In simulation we run it on exactly the same target-cohort sample sizes as every other method.
The strongest existing competitor is GSMR with HEIDI-outlier filtering (GSMR), applied to the transferred European instruments with target-ancestry LD supplied. It is the strongest because it already accounts for LD and sampling variance and already deletes inconsistent instruments. It is the closest thing to our canceller without the graded subtraction or the decision rule. Its reported European result for BMI on type 2 diabetes is a 3.29-fold risk increase per standard deviation of BMI. We found no reported cross-ancestry GSMR result on Lacuna. Its target-ancestry performance is therefore something we will measure, not quote.
We also include multi-ancestry fine-mapping restricted to fine-mapped variants as an instrument-selection baseline. HAUDI serves as a reference where PAGE individual-level data permit. Its reported gain is a 31.9% predictive R² improvement over ancestry-agnostic scores for white blood cell count in Hispanic/Latinos. That is a prediction metric, so it is a sanity check, not a head-to-head comparison.
All methods get the same summary statistics, reference panels, and instrument p-value thresholds. Any tunable threshold, such as the HEIDI cutoff, gets the same small grid search on held-out simulation replicates.
The key ablations isolate each ingredient. We run cancellation with a uniform prior versus an MFD prior versus a pilot-anchored prior. We run the threshold rule with $\rho$ predicted from features versus $\rho$ assumed at 0 or 1.
Beating GSMR means two things. Across simulated scenarios spanning LD mismatch, the rule's chosen estimator must have lower mean squared error than GSMR's. And the rule must choose the better of pooled versus target-only in more scenarios than either fixed strategy. Ties in the low-mismatch regime are expected and acceptable.
Why not just…?
Why not just run Mendelian randomization in the target ancestry alone? This is the safe half-duplex option, and the rule includes it as one branch. Its weakness is the small size of non-European cohorts: PAGE has roughly 17,300 African Americans and 11,776 Hispanic/Latinos, against up to 405,072 participants in GSMR's European data. Weak-instrument bias $W$ will therefore often dominate. The simulations directly measure the fraction of loci where it loses. If target-only wins almost everywhere, the threshold contribution is moot and we will report that.
Why not just restrict instruments to fine-mapped or functionally validated variants? This is the real threat. If every instrument were a validated causal variant, $\rho$ would be near 1 and no rule would be needed. The catch is supply: the file stresses that functional mapping of non-coding variants remains low-throughput and tied to particular cell contexts, so restriction discards most instruments and inflates variance. Our experiment compares "restrict to validated" against "cancel all, pool when $\rho > \rho^*$" at matched target sample sizes. If restriction matches us, we will recommend restriction and treat the pilots purely as calibration.
Why not just use an LD-mismatch score as a heuristic filter? Our $\rho$ predictor is partly built from such scores, so a reviewer could fairly ask whether the mismatch score alone suffices. The ablation that replaces predicted $\rho$ with raw mismatch rank, at a tuned cutoff, will settle it. Our added value must come from the pilot calibration and the explicit variance tradeoff in $\rho^*$, which a single-score filter ignores.
Proposed Experiments
The first experiment is semi-synthetic. We simulate phenotypes on PAGE haplotypes and European reference haplotypes, placing causal variants at positions drawn from functionally supported elements (CRISPRpath-type enhancers, FANTOM5). The European GWAS selects tagging instruments. We vary heritability, target sample size, and causal-position spread. Because $B$ and $\rho$ are known in simulation, we can check whether estimator error shows the predicted cliff at $\rho^*$. A smooth curve would falsify the threshold framing.
The second experiment is a positive control. Using white blood cell count with the Duffy-null region, the transferred European tags should show low $\rho$, and the rule should select target-anchored estimation. HAUDI's recovery of rs2814778 tells us the ground truth here.
The third experiment is a real-data replication. We transfer GSMR's European instruments for BMI to type 2 diabetes and LDL to type 2 diabetes into African American and Hispanic/Latino summary statistics. We report whether the rule pools or abstains and whether the direction of effect agrees with the European estimates.
The fourth experiment tests the rank-1 hypothesis. We measure what fraction of transfer bias lies along the leading eigenvector of the LD-difference matrix. If that fraction is high, cancelling only that direction should recover most of $\rho$, mirroring the rank-1 jamming result.
The decisive check across all four is whether $\rho$ predicted from features, calibrated only on pilot loci, generalizes to held-out pilot loci well enough that the rule's pooling decisions remain correct there. If calibration does not transfer across loci, the proposal reduces to the restriction baseline.
References
- When Full Duplex Wireless Meets Non-Orthogonal Multiple Access: Opportunities and Challenges
- Spectral and Energy Efficiencies in Full-Duplex Wireless Information and Power Transfer
- Improving Physical Layer Secrecy Using Full-Duplex Jamming Receivers
- Causal associations between risk factors and common diseases inferred from GWAS summary data
- An Efficient Lasso Framework for Admixture-Aware Polygenic Scores
- Improving polygenic risk prediction in admixed populations by explicitly modeling ancestral-differential effects via GAUDI
- Parallel characterization of cis-regulatory elements for multiple genes using CRISPRpath
- Metadata-guided feature disentanglement for functional genomics