Genetic studies of cause and effect, such as asking whether higher body mass index causes type-2 diabetes, increasingly use Mendelian randomization (MR). MR treats inherited DNA variants as natural experiments. A variant that raises an exposure is used to estimate that exposure's effect on a disease. Most of these variants were discovered in people of European ancestry. Usually they are not the biological cause themselves. They are "tags" that happen to be inherited together with the true causal variant, a correlation called linkage disequilibrium (LD). In other ancestries, especially African-ancestry populations, where these correlation blocks are shorter, a tag can stop carrying the signal. When an MR analysis is moved into such a population, some instruments quietly stop working, and nothing in the standard pipeline reports it. Wireless sensor networks deal with the same failure as a routine engineering problem. A gateway sends periodic "heartbeat" polls, so a node that went silent is detected before anyone relies on it. Network designers also deploy redundant nodes so that coverage survives when some nodes fail.
This note proposes carrying both habits into cross-ancestry MR. The first is a per-instrument heartbeat test, run on the target population's summary statistics. It labels each European-derived instrument as working, decayed through lost tagging, or behaving differently on the outcome, which suggests pleiotropy. The second is a way of choosing instruments that maximizes the instrument strength expected to survive the move into a new ancestry, rather than the strength observed in Europeans. The method for this second step is the dropout-aware clustering objective that Robust Marine Buoy Placement for Ship Detection Using Dropout K-Means developed for buoys that sink or are vandalized. The claim is narrow and testable. When target-ancestry samples are small or admixed, redundancy chosen for expected survival plus an explicit liveness check should give stronger instruments and better-calibrated causal estimates than European clumping, target-only selection, or post hoc outlier filters.
Core Claim
Moving an instrument set from a discovery ancestry to a target ancestry behaves like a network with unreliable links. Each true causal signal is a "node" that must stay reachable. Each tag variant is a "link" that survives or dies depending on whether the target population preserves its LD with the causal variant. Two things follow from this view, and current MR practice does neither.
First, instruments should be chosen so that every causal signal keeps at least one strong surviving tag in expectation. This differs from choosing one independent top variant per locus, which LD clumping does by design. Clumping removes exactly the backup tags that would carry the signal after the lead tag decays.
Second, each instrument's status in the target population should be checked actively before estimation, using a test that separates two failures. In LD decay, the exposure association falls and the outcome association falls with it. In a pleiotropic shift, the exposure association survives but the outcome-to-exposure ratio departs from the others. The research file names this separation as an open problem: MR-Egger and HEIDI "struggle to differentiate between genuine biological null effects, LD tagging decay, and horizontal pleiotropic masking."
How This Differs From Existing Work
The research file finds no paper joining sensor-network fault tracking with ancestry bias, so the relevant comparison is with the closest methods on each side.
On the genetics side, two papers motivate the gap. Causal associations between risk factors and common diseases inferred from GWAS summary data introduced GSMR, which builds LD into a generalized least-squares MR estimator and drops pleiotropic variants with the HEIDI-outlier test. Its authors state that their results depend on European data. HEIDI compares instruments against a single LD structure, so it cannot tell a variant that lost its tag from one with a different biology. Mendelian randomization study shows no causal relationship between psychiatric disorders and glaucoma in European and East Asian populations ran IVW, MR-Egger, weighted median and MR-PRESSO separately in European and East Asian biobanks. It treats each cohort as a static batch, and weakened instruments show up only as wider confidence intervals. We keep its parallel-cohort design but add the liveness check and the survival-aware instrument choice upstream.
Two papers constrain the design. Quantifying portable genetic effects and improving cross-ancestry genetic prediction with GWAS summary statistics (X-Wing) shows that portability is local: for mean corpuscular volume, about 1% of the genome carries over 60% of the European–East Asian genetic covariance. That result supports modeling survival per locus rather than as one global shrinkage factor. X-Wing is a prediction method that shrinks toward portable regions; it does not select MR instruments for redundancy or classify why an instrument failed. Evaluating Approaches for Constructing Polygenic Risk Scores for Prostate Cancer in Men of African and European Ancestry showed that 269 fine-mapped variants matched or beat genome-wide scores across ancestries. This is the most serious constraint on the proposal: if the causal variants were known, tagging would not decay and redundancy would add nothing. Our approach is therefore built around uncertain fine-mapping, where credible sets contain several candidate variants. It must be compared directly against fine-mapped instruments.
On the networking side, the dropout objective of Ng et al. supplies the method. Their formulation minimizes expected distance to the nearest surviving buoy under independent failures, and it can be solved exactly in $O(NK\log K)$ by sorting candidates for each demand point. On real vessel tracks with 30% dropout, ship-detection probability rose from 38% (classic k-means) to 52% (dropout k-median). k-Connectivity of Inhomogeneous Random Key Graphs With Unreliable Links shows that resilience is limited by the weakest node class. This motivates a design goal: a guaranteed minimum survival for the worst-covered causal signal, not only a high average. The gateway polling scheme in Design and implementation of Android phones-based home security system and the active-diagnosis category in A Survey on Proactive, Active and Passive Fault Diagnosis Protocols for WSNs supply the heartbeat idea: test for a response before relying on a component.
We take two things from the networking side and discard others. We do not import the analogy to cross-layer power boosting in A Cross-Layer Optimized Opportunistic Routing Scheme for Loss-and-Delay Sensitive WSNs. In MR, up-weighting degraded instruments plausibly amplifies pleiotropy instead of recovering signal, and nothing in the file suggests otherwise.
Research Program
Survival model. For each exposure, we start from European fine-mapping. Each association signal $j$ gets a credible set of candidate causal variants with posterior probabilities. For each candidate instrument $s$ and target ancestry $a$, we estimate a survival probability $q_{sa}$. This is the probability that $s$ keeps squared correlation above a threshold $\tau$ with the causal variant of signal $j$ in ancestry $a$, averaged over the credible-set posterior. The required LD matrices come from ancestry-specific panels such as TOP-LD, built on TOPMed whole genomes.
For admixed targets we average over local-ancestry proportions estimated with tools like XGMix. Using a mixture of panels in this way reflects the mosaic local-ancestry problem the research file identifies. Failures are correlated: tags in the same LD block in the same ancestry tend to fail together. We therefore treat survival events within a block as dependent and estimate their joint behavior by resampling across reference haplotypes, instead of assuming independence. Dropping the independence assumption is the main change from Ng et al.
Survival-aware selection. We choose an instrument set $S$ of size $K$, with selected tags allowed to be in LD with each other, to maximize expected surviving strength summed over signals and target ancestries:
$$\max_{S,\,|S|=K}\ \sum_{a}\pi_a \sum_{j} \mathbb{E}\Big[\max_{s\in S\cap C_j} w_{s}\,\mathbf{1}\{s \text{ survives in } a\}\Big],$$
Here $\pi_a$ weights target ancestries, $C_j$ is the set of candidate tags for signal $j$, and $w_s$ is the discovery-ancestry instrument strength of tag $s$. Under independent survival, the inner expectation has Ng et al.'s sorted closed form. Sort the selected tags by $w$; tag $s$ then contributes $w_s q_{sa}\prod_{t\text{ ranked above }s}(1-q_{ta})$. This makes greedy selection fast. With correlated survival we replace the closed form with Monte Carlo over the resampled failure patterns.
In the spirit of the weakest-class result, we also report a max-min variant that maximizes the worst signal's expected surviving strength. Because the selected tags are correlated, the MR estimator must account for LD. We therefore estimate with GSMR's generalized least squares, using target-ancestry LD.
Heartbeat test. For each selected instrument we predict its target-ancestry exposure association by projecting the fine-mapped causal effects through the target LD matrix. We compare this to the observed target exposure z-score. If the observed z-score falls short of a strong prediction, the instrument is flagged "unresponsive." If the exposure association matches the prediction but the outcome-to-exposure ratio deviates from the consensus of the other live instruments, it is flagged "behaving differently."
Only live instruments enter the estimator. The proportion of each failure type is reported as a diagnostic in its own right. Instrument selection uses only exposure summary statistics, which avoids selecting instruments on the outcome. The outcome is used only in the pleiotropy check, as HEIDI does.
Baselines
The naive reference point is the standard pipeline. European instruments are chosen by genome-wide significance and LD clumping, then applied unchanged in the target ancestry with inverse-variance-weighted (IVW) MR, as in the European arm of Zhang et al. This shows how much strength is lost silently. Any proposed method must beat it on retained instrument strength (mean F-statistic in the target) and on calibration in simulation.
The strongest competitor in well-powered settings is target-only selection. Instruments are chosen directly in the target GWAS (for example BioBank Japan) and analyzed with GSMR and HEIDI using target LD. It needs no transfer at all. To make the comparison fair, it receives the same target summary statistics, the same LD panels, and the same number of instruments $K$, and its significance threshold is tuned on the same held-out simulation grid.
The strongest competitor that handles cross-ancestry tagging is fine-mapped instrumenting in the style of Darst et al. It takes the single highest-posterior variant per credible set and drops redundancy. If our selection cannot beat it when credible sets are wide, the redundancy idea fails.
We also compare against European-selected instruments filtered with MR-PRESSO and MR-Egger in the target, which represents the "sensitivity analyses are enough" position.
Two ablations isolate the key ingredients. The first keeps survival-aware selection but assumes independent survival, which tests whether modeling correlated failure matters. The second applies the heartbeat filter to clumped instruments without redundant selection, which separates the value of checking from the value of backup tags.
For causal effects in East Asians there is no reported result on Lacuna for the target-ancestry instrument strength of transferred European instruments, so we must measure it ourselves. The nearest reported anchors come from Zhang et al. They report a European depression→glaucoma odds ratio of 1.15 (95% CI 0.93–1.42) and an East Asian odds ratio of 1.38 (p = 0.30). We will reproduce both before applying our method.
Why not just…?
Why not just select instruments in the target-ancestry GWAS? When the target GWAS is large, this is a real threat. With over 200,000 East Asian participants, as in BioBank Japan and China Kadoorie Biobank combined in Zhang et al., target-only selection may match or beat any transfer method. We expect our advantage only where target data are small or admixed. The sample-size sweep described below is designed to find where, if anywhere, the curves cross. If target-only selection wins at every sample size, the selection half of the proposal fails, and only the heartbeat diagnostic remains worth publishing.
Why not just use fine-mapped causal variants? If fine-mapping pinpointed causal variants, LD decay would be irrelevant, and Darst et al. show that compact fine-mapped sets transfer well. In practice credible sets often contain many variants, and the lead variant can itself be a tag. We stratify every comparison by credible-set size. The prediction is that survival-aware redundancy helps only for signals with large credible sets and adds nothing for well-resolved ones. A uniform gain at all set sizes would suggest an artifact.
Why not just rely on MR-PRESSO, MR-Egger or HEIDI? These tests look for instruments whose ratio estimates disagree with the rest. Zhu et al.'s HEIDI was calibrated against a single European LD structure. Zhang et al. ran the whole suite of sensitivity analyses and still could not exclude unmeasured pleiotropy. None of these tests uses the predicted target exposure association, so none can tell lost relevance apart from pleiotropy. Simulations where the failure type is planted directly will show whether the heartbeat test classifies failures better than an outlier flag does.
Why not just apply a multi-ancestry prediction method like X-Wing? X-Wing improves polygenic prediction, with a median relative R² gain of 35.2% in East Asians across 31 traits reported in Miao et al. Its portability annotations could replace our survival probabilities, and we will test that substitution as an ablation. X-Wing does not, however, select a small set of individually checkable instruments of the kind MR's assumptions require.
Proposed Experiments
The first experiment is a semi-synthetic simulation built on real LD. We simulate exposures and outcomes from TOPMed-based LD for European, East Asian, African-ancestry, and admixed Hispanic/Latino groups. In each simulation we plant known causal variants, known causal effects, and ancestry-specific pleiotropy in a controlled fraction of loci. The metrics are bias, 95% interval coverage, and power for the causal effect. We also measure retained target F-statistics, and the heartbeat test's accuracy at separating "unresponsive" from "behaving differently" instruments against the planted ground truth. The decisive check is whether survival-aware selection improves coverage over clumping in African-ancestry and admixed targets without inflating false positives under the null.
The second experiment uses real summary statistics from European discovery data (UK Biobank, FinnGen, Psychiatric Genomics Consortium exposures) and East Asian targets (BioBank Japan, China Kadoorie Biobank). We subsample the target cohorts to mimic underpowered ancestries.
Positive controls come from Zhu et al.'s European findings, such as body mass index on type-2 diabetes (a 3.29-fold risk per standard deviation) and LDL cholesterol on type-2 diabetes. Each method should recover the direction of these effects in East Asians, and with fewer target samples than the alternatives need. The null controls are Zhang et al.'s psychiatric→glaucoma pairs, where each method should stay null.
The third experiment applies the pipeline to multi-ethnic PAGE data, to test admixed targets where discrete LD panels are known to fit poorly.
The proposal is decisively supported if survival-aware selection plus heartbeat filtering beats fine-mapped instruments on coverage for signals with wide credible sets, and beats target-only selection below some target sample size. It is refuted if fine-mapped single instruments match it at every credible-set size. It is also refuted if the heartbeat test classifies planted failures no better than MR-PRESSO's outlier flag.
References
- Robust Marine Buoy Placement for Ship Detection Using Dropout K-Means
- 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
- Quantifying portable genetic effects and improving cross-ancestry genetic prediction with GWAS summary statistics
- Evaluating approaches for constructing polygenic risk scores for prostate cancer in men of African and European ancestry
- $k$ -Connectivity of Inhomogeneous Random Key Graphs With Unreliable Links
- Design and implementation of Android phones-based home security system
- A Survey on Proactive, Active and Passive Fault Diagnosis Protocols for WSNs: Network Operation Perspective
- A Cross-Layer Optimized Opportunistic Routing Scheme for Loss-and-Delay Sensitive WSNs
- TOP-LD: A tool to explore linkage disequilibrium with TOPMed whole-genome sequence data
- XGMix: Local-Ancestry Inference with Stacked XGBoost