Transcription Factor Activity as a Carrier of Missing Trans Heritability

Much of the genetic variance in complex traits that twin studies detect is never found by genome-wide association studies (GWAS), which test one DNA variant at a time. A growing view holds that a large share of this "missing heritability" acts in trans: a variant changes a gene far away by altering a regulator, which then shifts many downstream genes. Trans effects are hard to find because each one is small. A search over every variant–gene pair also carries an enormous multiple-testing burden. This proposal borrows one idea from gene regulatory network inference: the amount of a transcription factor's (TF's) mRNA is a poor readout of how active that factor is. A better readout is the coordinated shift of the TF's target genes, its regulon.

We propose to treat this regulon-derived activity as a molecular trait in its own right. The plan is to map the genetic variants that predict a TF's activity, which we call activity-QTLs (quantitative trait loci, variants associated with a measurable quantity). Variants sitting next to the target genes are excluded, so the signal cannot leak in from ordinary local effects. We then ask how much complex-trait heritability these variants carry beyond standard local expression models. The concrete payoff is a test with roughly 1,600 hypotheses, one per TF, that targets trans regulation. The scientific question is whether the "master regulator" bottlenecks of network biology are where polygenic variance actually sits.

Core Claim

Genotype-predicted TF activity, built from regulon targets with every target's own neighborhood removed from the predictor, captures trait heritability that cis-only transcriptome-wide association studies (TWAS) miss. It also captures heritability that mediator-mRNA models such as MOSTWAS miss. The reason is that activity sums many small, same-direction trans shifts into one measurable quantity.

The averaging argument is simple. Suppose a variant shifts each of $|R|$ targets by $\delta$ standard deviations and each target has independent noise. The summed score then moves by $\delta\sqrt{|R|}$ relative to its noise, because $|R|$ coherent shifts add up linearly while the noise grows only as $\sqrt{|R|}$. If target noise has average pairwise correlation $\rho$, the gain levels off near $1/\sqrt{\rho}$. This is why controlling for hidden factors matters so much, and why a matched random-gene-set control is the central check.

The claim is new relative to the work it builds on. Califano et al.'s integrative network association studies reconstruct regulons and infer master-regulator activity with MARINa. They then test whether GWAS variants are enriched in a subnetwork (an 8-fold enrichment for type 2 diabetes), but they never build a genotype-to-activity predictor or partition heritability. MOSTWAS reaches distal variants through local mediators. It uses the mediators' mRNA (or methylation and microRNA) as inputs to predict each target gene separately, so it cannot see variants that change a TF's activity without changing its transcript: variants in cofactors, modifying enzymes, or upstream signaling. The bipartite eQTL network study shows heritability concentrates in "local hub" variants of SNP–gene modules. Its modules, however, are statistical eQTL communities, not TF-to-target regulons, and contain no activity estimate. We reuse its partitioning method, stratified LD score regression (S-LDSC, which apportions heritability among annotated variant classes while accounting for correlation between nearby variants). We also use its local hubs as an external point of comparison.

Research Program

Step 1: regulons. We first build regulons per tissue. geneRNIB and CausalBench both find that simple expression methods (Pearson correlation, GRNBoost2) match or beat prior-heavy pipelines and foundation models. Taking those benchmarks seriously, we do not bet on a sophisticated network. The network source becomes an experimental factor with three levels:

  1. GRNBoost2 restricted to TF regulators;
  2. curated and ChIP-seq prior regulons of the kind compiled by metaTF;
  3. the scFM-derived edges of UGRN where they can be computed.

Step 2: activity scores. For TF $t$ with regulon $R_t$, sample $i$ gets the activity score

$$a_{t,i} = \frac{1}{\sqrt{|R_t|}} \sum_{g \in R_t} s_{tg}\, z_{g,i},$$

where $z_{g,i}$ is the expression of target $g$ in sample $i$ after removing hidden technical factors, and $s_{tg}$ is the sign of the inferred edge. We also run a VIPER-style enrichment version for comparison.

Step 3: activity-QTLs. We then fit a sparse genotype model for $a_{t,i}$. The allowed variants exclude a 1 Mb window around every target in $R_t$, with an additional LD-pruning buffer. Variants near the TF itself are allowed, so an activity-QTL can be a local TF variant, a distal variant, or anything in between, but never a local variant of a target.

Step 4: trait tests. The predicted activities are tested against GWAS summary statistics in the usual TWAS way. Activity-QTL variants also become an S-LDSC annotation, fitted jointly with the baseline annotations, a cis-eQTL annotation, and a MOSTWAS distal-eQTL annotation.

Settings.

Purpose Data
Main heritability analyses GTEx v8 tissues (GTEx)
Multi-omic comparison with MOSTWAS ROS/MAP brain and TCGA-BRCA, the cohorts where MOSTWAS reported results
Trait summary statistics Alzheimer's disease (IGAP), major depression (PGC), breast cancer (iCOGs), and UK Biobank traits including C-reactive protein

Choosing whole blood versus a solid tissue matters. The multilayer transcriptome study found blood communities to be poor proxies for other tissues, so each trait is paired with its relevant tissue.

Baselines

Do-nothing reference. S-LDSC with only the standard baseline annotations and a cis-eQTL annotation. This is the heritability already explained before any network enters. Activity-QTLs must add a significant enrichment and a non-trivial heritability share conditional on these annotations.

Cis-only TWAS. An elastic net over variants within 1 Mb of each gene, including the TF's own mRNA. This is the default practice and the obvious skeptic's tool. Beating it means two things:

  • predicted activity identifies trait-associated TFs whose cis-predicted mRNA is not associated;
  • activity-QTL heritability survives conditioning on the cis-eQTL annotation.

MOSTWAS (strongest existing method). MOSTWAS is the closest competitor: it already captures distal variation through regulators. It reports additive increases of 1–2% in expression variance explained over local-only models, with 86% (MeTWAS) and 79% (DePMA) of models showing higher cross-validated $R^2$ than local-only, in ROS/MAP and TCGA-BRCA.

We run MOSTWAS on the same cohorts and the same target genes, with matched hidden-factor correction and the same nested cross-validation folds and hyperparameter grid size. Beating it has two parts:

  • Expression level. Adding predicted upstream-TF activity as a covariate to MeTWAS raises held-out target-gene $R^2$ in a majority of genes, by an amount comparable to MOSTWAS's own 1–2% gain over local-only models.
  • Trait level. The activity-QTL annotation retains heritability enrichment after the MOSTWAS distal-eQTL annotation is in the model.

Ablation 1: random gene sets. This isolates the key ingredient. Each TF's regulon is replaced by a random gene set matched in size, mean expression, and pairwise-correlation distribution, and the whole pipeline is rerun. If random sets yield activity-QTLs and heritability at similar rates, the signal is generic coherence, such as residual technical factors, not regulation.

Ablation 2: co-expression modules. Each regulon is replaced by an undirected co-expression module (Louvain communities as in the multilayer study) and summarized by its first principal component. This is the alternative that could plausibly match us.

Ablation 3: no cis-target exclusion. We drop the cis-target exclusion and show how much apparent signal is merely local eQTLs of targets leaking into the score.

Why not just…?

Why not just run cis-TWAS on the TF's own mRNA? That model only sees variants that change TF transcript levels. A central finding of the network literature is that activity often moves without transcript change, for example SOX10 holding subtype circuits in glioblastoma (Wu et al.), and C/EBP and STAT3 identified from regulon activity in Califano et al.. The experiment that settles this is direct: if activity-QTL heritability and TF-level associations collapse onto the TF's own cis-eQTLs, activity adds nothing and we report that.

Why not just extend MOSTWAS? MOSTWAS is the real threat, and it is run head-to-head. Its mediators are measured mRNA or methylation levels fed into one target-gene model at a time, with permutation-based mediation tests that its authors note are computationally heavy. Our test is one model per TF instead of a mediation test per gene–mediator pair. It also targets variants that change regulator function rather than regulator abundance. If adding activity to MeTWAS does not raise held-out $R^2$, and the activity annotation loses enrichment once MOSTWAS distal eQTLs are included, the idea fails.

Why not just use co-expression modules instead of TF regulons? This is the alternative most likely to match us. geneRNIB shows Pearson correlation is a hard baseline for network recovery, and the multilayer study raised the TWAS true-positive rate from $\pi_1 = 0.37$ to $0.45$ using undirected communities alone. That is why the default regulon is itself expression-derived (GRNBoost2). What we add is anchoring each gene set on a TF regulator with signed edges, which yields a named upstream hypothesis that can be checked by perturbation. If module eigengenes match regulon activity on heritability capture, we will say so. The contribution then narrows to a check that the TF anchor identifies the right regulator, which the perturbation test below can confirm or refute.

Why not just use a genome-wide PRS or gene-set enrichment over GO/KEGG? A PRS does not say which regulator carries the variance. Canonical pathways missed the type 2 diabetes signal that de novo networks found in Califano et al.. We include GO/KEGG gene-set annotations in the S-LDSC model as a further conditioning set.

Proposed Experiments and Decisive Checks

Check 1: does the activity score track the TF at all? Before any genetics, we validate the readout on perturbation data where the true regulator is known:

  • the Replogle, Norman, and Adamson CRISPR screens in geneRNIB;
  • the K562 and RPE1 sets of CausalBench;
  • the 40-TF CRISPRi experiment in human embryonic stem cells used by metaTF.

The metric is how often a knocked-down TF's activity score ranks first among all TFs' scores in the perturbed cells, compared across the three network sources and against the TF's own mRNA. No reported result was found on Lacuna for this exact ranking metric, so all numbers are ours. This check decides which network source enters the genetic analysis. Because benchmark results favor simple methods, we expect GRNBoost2 regulons to perform as well as the others, and we will use whichever source wins.

Check 2: leakage audit. Ablation 3, rerun with and without the exclusion window, reports the fraction of activity-QTLs that fall within LD of any target's cis region.

Check 3: heritability with joint conditioning. S-LDSC across GTEx tissue–trait pairs reports the activity-QTL annotation's enrichment and heritability share. The model is conditioned jointly on the baseline, cis-eQTL, MOSTWAS distal-eQTL, co-expression-module, and random-set annotations. We also measure overlap with the local-hub variants of the bipartite eQTL network study. Activity-QTLs that coincide with those hubs would suggest the hubs are partly TF-activity variants.

Check 4: replication of TF associations. Activity-WAS hits from one GWAS must replicate in an independent GWAS of the same trait where one exists, for example breast cancer in iCOGs versus UK Biobank. Top TFs whose regulons are perturbed in the screens above are checked for whether their perturbation signature agrees in sign with the predicted trait-increasing direction.

What counts as a positive result. The proposal succeeds if all of the following hold:

  • activity-QTL heritability enrichment survives every conditioning set in the joint S-LDSC model;
  • it clearly exceeds the matched random-set control;
  • it improves on MOSTWAS on held-out expression $R^2$.

If only the random-set and co-expression controls fail to separate from regulon activity, the honest conclusion is that coherent trans signal matters but directed regulation does not. That would itself be a useful answer to the open question in geneRNIB, carried over to heritability.