Do Overexpression-Born Protein Interactions Blur Where Heritability Lives?

Many methods that try to recover "missing heritability" lean on maps of which proteins physically bind which. Missing heritability is the gap between how much of a trait family studies attribute to genes and how much the individually significant variants from genome-wide scans can explain. These methods use the protein-binding map to group weak genetic signals into modules of genes that work together. Much of that map, however, was built by flooding cells with a single protein far above its natural level. At those concentrations, mass action alone forces low-affinity contacts that never happen in a normal cell. This note proposes to label every edge in a standard interaction map by the dosage regime of the experiment that found it. We would then test whether edges seen only under forced overexpression systematically pull genetic signal toward the wrong genes and smear the module boundaries that heritability analyses depend on.

The claim is concrete and falsifiable. Suppose we remove interactions supported only by overexpression assays, and remove nothing else. If heritability then concentrates into fewer, more trait-specific modules, beyond what a matched random removal achieves, then the biophysics of overexpression is a measurable source of error in statistical genetics. If it does not, network-based heritability methods are robust to this contamination, which is also worth knowing.

Core Claim

Two literatures sit side by side without touching. The first shows that non-physiological expression creates interactions and localizations that are not native. A scalable strategy for high-throughput GFP tagging of endogenous human proteins built an endogenous split-GFP tagging pipeline because plasmid-driven expression produced mislocalization and aberrant binding. Multiplexed selectivity screening of anti-GPCR antibodies found that high receptor concentrations inflate off-target binding. Prevalent Presence of Periodic Actin-spectrin-based Membrane Skeleton reports that overexpressed tagged structural proteins can assemble lattices where none exist natively. These papers motivate the hypothesis: they establish the mechanism by which false edges arise.

The second literature uses networks to localize heritability. MPRS-NetWAS propagates Alzheimer's GWAS signal over a general protein-interaction network with Hierarchical HotNet and scores module-level risk. The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution shows that heritability concentrates in a minority of tissue-specific network modules and their local hubs. For example, breast cancer SNPs fall into only 12.6% of skin-network modules, and about 29% of modules are trait-specific. These papers supply the methods and the measuring stick. Neither asks how the provenance of its network edges shapes its answer. The research file's own synthesis notes that no paper connects the two literatures directly.

Our specific prediction follows from mass action. Overexpression-only edges should preferentially attach to abundant, aggregation-prone, or "sticky" proteins. This would inflate their connectivity and make them false hubs that absorb propagated GWAS signal. The consequence is that modules grow larger and less trait-specific, and heritability looks more diffuse, more "omnigenic," than it is.

Research Program

The first step is to annotate edges. For each interaction in a widely used physical interaction database, we record the experimental systems that support it. We then sort edges into three groups. "Forced-dosage only" edges are supported exclusively by yeast two-hybrid or by affinity purification with an overexpressed bait. "Native-dosage" edges are supported by at least one assay at endogenous levels, such as endogenous-tag pulldown, co-fractionation, or co-immunoprecipitation of untagged proteins. The third group is mixed. We then characterize the forced-dosage-only edges. We test whether their endpoints are enriched for high abundance, for disordered or phase-separating regions (the concentration-sensitive behaviour UBQLN2 work highlights), and for inflated degree.

The second step builds parallel networks and reruns the heritability analyses unchanged on each. There are four versions: the full network, a native-dosage-only network, a version that keeps edges weighted down by a factor learned from replication across assay types, and matched random-removal controls. On each network we rerun Hierarchical HotNet propagation as in MPRS-NetWAS. We then convert the resulting modules into SNP annotations, using a window around module genes, and estimate heritability enrichment with stratified LD score regression. This is the method Márquez-Luna and colleagues (The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution) used, which asks what fraction of heritability falls in SNPs inside each annotation relative to the fraction of SNPs it contains. The traits are those whose summary statistics the file describes. These are the four Global Lipids Genetics Consortium lipid traits (about 188,577 individuals, as used by Brandes et al.), Alzheimer's disease from IGAP and Jansen et al., and UK Biobank traits from the phenome-wide resource of Phenome-wide heritability analysis of the UK Biobank.

The third step is a small wet-lab test of the edges that matter most. We take the handful of forced-dosage-only edges whose removal moves a GWAS-implicated gene into or out of a top module. For each, we tag both partners at their endogenous loci with split GFP in HEK293T cells, following the endogenous tagging pipeline. We then test co-localization and pulldown at native levels, and compare against the same pair under CMV-driven overexpression. That paper reports a 63% single-guide tagging rate and roughly 30% proteome detectability with one tag, or about 60% with four tandem tags. So we will preselect pairs whose proteins are abundant enough to detect, and we accept that low-abundance pairs cannot be adjudicated this way.

Baselines

The naive reference point is the unmodified pipeline. This means MPRS-NetWAS-style propagation and module scoring on the full general interaction network, followed by stratified LD score regression on the resulting modules. It is the strongest existing method for this question because it is exactly what current network-based heritability work does. We run it with the same summary statistics, LD reference, gene-window size, and HotNet settings as every variant. No reported result found on Lacuna for module-level heritability enrichment or variance explained by MPRS-NetWAS, since the file notes the paper reports four Alzheimer's modules but no numeric heritability values. We will therefore reproduce it ourselves and treat our reproduction as the reference.

The strongest non-protein competitor is the eQTL bipartite-network approach of Márquez-Luna et al., built from GTEx v8 across 29 tissues with community detection. It needs no protein-interaction database at all. Its reported concentration figures (12.6% of modules carrying breast cancer SNPs in skin; about 29% trait-specific modules) are the targets a cleaned protein network must approach or exceed. We will recompute them under our trait set and LD settings so the comparison is not across pipelines.

The ablation that isolates our key ingredient is matched random removal. From the full network we delete random edge sets that equal the forced-dosage-only set in number and in the degree distribution of their endpoints. We also run a second random set matched on database confidence score. If removing forced-dosage-only edges concentrates heritability no more than these controls do, the effect is just sparsification and our hypothesis fails.

We count a win if two things hold. First, the native-dosage network must yield modules capturing a larger share of heritability per share of SNPs than both the full network and the 95th percentile of matched random removals, across most traits. Second, the native-dosage network must yield more trait-specific modules, with no loss in held-out Alzheimer's stage discrimination when module risk scores are computed in ADNI.

Why not just…?

Why not just threshold the database confidence score? Confidence scores mostly reward replication. An interaction that is reproducible under overexpression is still reproducible, so thresholding can keep exactly the artefacts we target. This is a real threat, because high-confidence filtering might remove most forced-dosage-only edges anyway. Our confidence-matched random control settles it. If filtering by dosage regime beats filtering by an equal number of low-confidence edges, provenance carries information that confidence does not. If not, we will report that confidence filtering suffices.

Why not just drop protein interactions and use eQTL networks? Regulatory networks already localize heritability well, as Márquez-Luna et al. show. But they capture who regulates whose transcript, not which proteins form complexes. Variants acting through protein stability, which the folliculin degradation study shows can leave mRNA untouched, are invisible to them. We compare head to head. If the eQTL network matches the cleaned protein network on every trait, the protein layer adds nothing and the practical recommendation is to drop it. We will also test whether combining the two improves on either alone.

Why not just use gene-set enrichment without networks? Pathway and gene-set aggregation, as in the warfarin and rare-variant work summarized in the file, avoids network artefacts entirely. Pathway annotations, though, are partly built from the same interaction evidence, and they cannot place unannotated genes. We include curated-pathway annotations as an extra comparison in stratified LD score regression. If they equal the cleaned network, our contribution narrows to a diagnosis of existing network methods rather than a better replacement.

Proposed Experiments and Decisive Checks

The decisive experiment is the comparison between the native-dosage network and the matched random removals, measured by heritability enrichment of the top modules across lipid, Alzheimer's, and UK Biobank traits.

A second check guards against the most worrying confound, literature bias. Native-dosage assays are concentrated on well-studied genes, which overlap with known GWAS hits. A cleaned network could therefore look better simply by encoding prior GWAS knowledge. To control this, we derive modules with signal from one trait and evaluate enrichment on unrelated traits. We also repeat the analysis after removing all genes within a fixed distance of previously reported genome-wide significant loci. We further check that forced-dosage-only edges do not simply mark the understudied genes that yeast two-hybrid screens sample more evenly. To do this, we stratify results by publication count per gene.

A third check is mechanistic. If our account is right, the false hubs absorbing signal in the full network should be disproportionately abundant or disordered proteins. In the wet-lab arm, their contested edges should mostly fail to appear at endogenous dosage while appearing under CMV overexpression.

The fourth check uses Alzheimer's disease as a clinical anchor. We recompute module risk scores in ADNI from cleaned and uncleaned modules and compare their discrimination of the five clinical stages and of conversion from mild cognitive impairment. The goal is to see whether cleaner modules are also more predictive or merely tidier.

A negative result on all four checks would be informative too. It would tell the field that network-based heritability partitioning is robust to the best-documented artefact in its input data.