Pleiotropy is the situation where one genetic variant affects several diseases. Researchers increasingly explore it through interactive "disease-disease networks": each disease is a node, and two diseases are joined when they share a genetic association. These maps are persuasive because hubs and clusters jump out at the eye. That is also the risk. The edges in today's tools come from one fixed rule, usually "both diseases have a significant association at the same variant." Neighbouring variants on a chromosome are inherited together, a correlation called linkage disequilibrium (LD). Because of LD, two diseases driven by different nearby causal variants can look like they share one. A hub the user sees may be a real shared mechanism, or it may come from how the edges were defined.
We propose a network view in which every edge carries its own evidence. For each disease pair we compute sharing under several edge definitions of increasing statistical care: single-variant thresholding, LD-pruned thresholding, and LD-aware local genetic correlation scanning. We also recompute each edge on independent halves of the data. The interface then shows how stable each edge and hub is across these definitions and data splits, and the user can filter the network by that stability. The claim is testable in two ways. Statistically, stability-filtered edges should replicate in independent data more often than edges from any single rule. For users, analysts working with the stability view should nominate pleiotropic hubs that replicate more often than analysts using existing tools.
Core Claim
The topology of a pleiotropy network is not a fixed property of the data. Hub identity, cluster membership and even the sign of sharing change depending on how an edge is defined. Making that dependence visible improves the accuracy of what people conclude from the network. The contribution is therefore an edge representation, plus an evaluation of whether people use it well, and not simply a new dashboard. The sign matters because local correlation scanning separates sharing in the same direction from sharing in opposite directions. A single genome-wide correlation cannot do this, since opposite local effects cancel out.
Relation to Prior Work
The work closest to ours is NETMAGE. It turns UK Biobank PheWAS summary statistics into an interactive sigma.js disease network. Its edges come from shared single variants passing p < 1e-8, minor allele frequency > 0.01 and more than 200 cases, giving 260 phenotypes and 2,642 associations. This work motivates ours: its edge rule is exactly the LD-blind definition we want to stress-test. We also reuse its data and its rendering approach. NETMAGE shows one network. We show how far that network can be trusted, edge by edge.
LOGODetect supplies our main statistical method. It slides windows of varying size along the genome and tests whether two traits' association scores correlate locally, after adjusting for LD. Across seven neuropsychiatric traits it found 227 shared segments and several hubs affecting five or more traits. In the paper's discovery-to-UK-Biobank replication analysis, 17 of 24 regions replicated. Its outputs are static tables and mirror Manhattan plots, and it was not used to build networks. We run it per disease pair as one edge definition and reuse its discovery/replication design as our ground truth.
Two systems show that coordinated interactive views and natural-language querying are feasible for genetics, but neither questions whether the associations it displays are robust. GenAMap provides linked views over multi-task sparse regression and notes that such models struggle to give analytical p-values. GENEVIC covers only three phenotypes using precomputed polygenic-score weights. POGO builds multi-layer comorbidity networks from curated databases, with static output.
We read the GenomicSEM result in the ADHD/autism polygenic score study as a cautionary constraint. There, a conventional autism score was associated with six unrelated disorders and the refined score with none. Apparent cross-trait sharing can be heavily inflated, which is exactly the inflation our edge annotations aim to expose.
Research Program
The computational layer has three steps. First, it computes edges for every pair of the 260 NETMAGE phenotypes under four definitions: the original NETMAGE rule, the same rule after LD-based pruning within a 1 Mb window, LOGODetect local correlation segments with their sign, and LOGODetect run at several settings of its adaptive tuning parameter θ.
Second, it computes stability. Each pair is re-estimated on split halves where individual-level splits are available, and on independent cohorts otherwise. For the seven neuropsychiatric traits, the discovery consortia serve as one source and UK Biobank as the other, following LOGODetect. Each edge then gets two stability scores. Definition stability is the fraction of the four definitions that recover the edge. Data stability is the fraction of resamples that recover it.
Third, it marks hub stability. Each node's degree is recomputed under every definition, and nodes whose hub rank collapses across definitions are flagged.
The interface layer extends NETMAGE-style rendering. Edge opacity encodes stability, and edge colour encodes the sign of local correlation. A threshold slider re-renders the network live. A "disagreement" view highlights edges present under single-variant rules but absent under LD-aware scanning. Clicking an edge opens its local correlation track along the chromosome. Heavy computation is done in advance, so interaction needs only lookups. This avoids GenAMap's dependence on a compute cluster and the latency of language-model repair loops.
Baselines
The first naive reference point is a static figure set: a PheWAS Manhattan plot, a genome-wide genetic correlation heatmap, and LOGODetect-style mirror Manhattan plots, all given to users as an image packet. This is what most papers ship today.
The second reference point is tabular filtering. Users get the summary statistics in a queryable table, are allowed SQL or Python, and are given the same time budget.
The strongest existing interactive competitor is NETMAGE, run exactly as published on the same UK Biobank data and thresholds. We know of no replication rate reported for its edges: no reported result found on Lacuna for NETMAGE edge replication. We will therefore measure it ourselves. The strongest statistical competitor is LOGODetect alone. Its reported 17/24 regional replication, about 71%, in Detecting local genetic correlations with scan statistics is the number a stability-filtered edge set must match or exceed at comparable edge counts.
The key ablation is our full interface with the stability encoding removed, so that only LOGODetect edges are drawn at a single θ. This separates the value of showing robustness from the value of simply using a better edge rule. A second ablation keeps stability across data splits but drops stability across definitions.
All conditions use the same summary statistics, the same LD reference, the same task set and the same time budget. Every statistical rule's threshold is tuned on the discovery split only. We count our method as beating a baseline only if it wins at matched numbers of edges, because the precision of a smaller edge set is trivially higher.
Why not just…?
Why not just LD-prune variants before drawing NETMAGE edges? This is the most serious threat. Pruning removes redundant correlated variants within one trait, but it does not test whether two traits share a causal signal or merely have different causal variants nearby. It may still remove most spurious edges. Our edge-replication experiment includes pruned NETMAGE as an explicit arm. If it matches stability filtering in replication precision at matched edge counts, the statistical half of our claim fails. The contribution would then shrink to the interface finding, and we would say so.
Why not just use genome-wide genetic correlation between disease pairs? A single correlation coefficient cannot localise sharing. It also misses pairs whose local effects run in opposite directions. LOGODetect reports outperforming coloc, gwas-pw and ρ-HESS in simulations, and it found hubs that a global summary hides. Our replication experiment includes a correlation-thresholded network as a further arm, so this can be checked directly rather than assumed.
Why not just ship LOGODetect's static plots? Static figures may be enough for experts studying one pair of traits. Our user study tests whether that holds at the scale of 260 phenotypes, where a static packet has tens of thousands of possible pairs. The static-packet condition is exactly this alternative. If expert users perform equally well with it, the interactive layer is unnecessary for experts, and we would report that.
Proposed Experiments
The first experiment tests whether edges replicate. Using the seven neuropsychiatric traits, plus UK Biobank phenotype pairs where split-sample statistics can be built, we measure each edge set's replication precision against held-out data, at matched edge counts. The edge sets are NETMAGE, pruned NETMAGE, correlation-thresholded, LOGODetect at a single θ, and stability-filtered. Separately, we report the fraction of NETMAGE's 2,642 associations that become disagreement edges under LD-aware scanning, along with the hubs whose rank changes. The decisive check is that stability filtering beats both pruned NETMAGE and single-θ LOGODetect in replication precision. LOGODetect's single setting is anchored by its reported 17/24.
The second experiment tests whether people reach correct conclusions. Statistical geneticists and biologists without programming backgrounds complete tasks such as "nominate three diseases that share a genetic hub with schizophrenia" and "is this edge a shared mechanism?" Each participant works under the static, tabular, NETMAGE, ablated and full conditions in a counterbalanced order. Ground truth is replication in held-out data. The outcomes are the replication precision of nominated hubs, the error rate on the disagreement edges flagged in the first experiment, time per task, and confidence calibration. The decisive check is that the full interface beats the ablation. Without that result, any gain could be credited to LOGODetect alone, not to showing robustness.
References
- NETMAGE: a humaN-disEase phenoType MAp GEnerator for the Visualization of PheWAS
- Detecting local genetic correlations with scan statistics
- GWAS in a Box: Statistical and Visual Analytics of Structured Associations via GenAMap
- GENEVIC: GENetic data Exploration and Visualization via Intelligent interactive Console
- How to build personalized multi-omics comorbidity profiles
- Enhancing the discriminatory power of polygenic scores for ADHD and autism in clinical and non-clinical samples