Most of the genetic risk for common diseases lies in many small-effect variants scattered through the non-coding genome. Genome-wide association studies (GWAS) find only a fraction of it, and the rest is called "missing heritability." A natural guess is that these variants act by changing how genes are switched on and off in a specific cell at a specific moment. Current tools that split heritability across the genome describe that switching as a static wiring diagram, and they measure it in ground-up tissue that averages many cell types together. This note proposes a narrower test with a clear failure condition. In CD4+ T cells, the regulatory proteins called transcription factors change which genes they control as the cell passes through successive divisions after activation. We ask whether disease heritability is concentrated in the DNA that mediates these changing connections, and whether that concentration goes beyond what is already explained by the genes whose expression levels change.
In two sentences for a colleague: build a separate map of transcription-factor-to-gene connections for each T-cell division stage, then find the connections that switch between stages. Then test with standard heritability partitioning whether immune-disease genetic risk sits near those switching connections more than near genes that merely rise or fall in expression.
Core Claim
Our hypothesis is that heritability for immune-mediated diseases is enriched at regulatory sites belonging to transcription-factor–target connections that rewire during early T-cell division. We also expect this enrichment to survive conditioning on expression-change annotations and on static network-hub annotations. The claim is new in a specific way. It moves the unit of heritability partitioning from nodes (genes, or modules of a fixed network) to edges that change between measured cell states. It also defines those states by division count rather than by clock time or by averaged tissue.
The claim is built from three pieces of prior work. The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution motivates it. That paper grouped bulk-tissue GTEx v8 variant-to-gene networks into modules and used stratified LD score regression (S-LDSC), a regression of GWAS association strength on how much each variant is correlated with variants in an annotation, to show that heritability concentrates at "local hubs" inside tissue-specific modules. Its networks are fixed in time and come from homogenized tissue, and the authors flag both as limitations. Division state reveals hidden genetic regulation during T cell activation and identifies immune disease–linked gene programmes supplies the temporal axis. It sorted CD4+ T cells into six division states and trained CellDivider, a small neural network, to predict division state from a single cell's transcriptome. It then found that immune disease risk variants act mainly through gene programs of early division states. Its analysis stays at the level of genes and programs and never asks about connections. Estimating drivers of cell state transitions using gene regulatory network models supplies the rewiring method. Its MONSTER approach models the network at a later state as a linear transformation of the earlier one and scores each transcription factor by how much of its regulatory weight moves to new targets. It also supplies an important constraint: naive edge-by-edge subtraction replicated with $R^2 < 0.01$ across cohorts, while factor-level scores replicated (7 of the top 10 factors shared between two blood cohorts). The proposal is designed around that negative result.
Research Program
We first build one regulatory network per division state. Using the division-sorted single-cell data from Soskic et al., we aggregate cells into pseudo-bulk profiles per donor and per division state, which keeps donor as the unit of replication. We then fit bipartite networks of transcription factors to target genes in the PANDA style used by Lopes‐Ramos et al.. This style combines a motif prior, known protein–protein interactions and co-expression, and it has already shown that factors can release control of target genes without any change in their own mRNA. For cells without a division label, such as cells from other datasets, CellDivider supplies the state. Its reported test performance is a Pearson correlation of 0.881 and a mean absolute error of 0.428 division units on 25,694 held-out cells, so we treat its labels as usable but noisy and propagate that uncertainty.
Second, we quantify rewiring without trusting single edges. Between consecutive division states we fit MONSTER transition matrices and compute each factor's differential involvement score, with null distributions from permuting division labels within donors. From factors that rewire significantly, we define edges as gained or lost only when the change holds across donor-split bootstraps. We add the edge-level stability requirement because the MONSTER result shows raw edge differences are mostly noise.
Third, we map edges to DNA. For each retained edge, the annotation contains variants that fall in a motif for the rewiring factor inside the target gene's promoter or linked open-chromatin region. A broader version uses a gene window around the target. Matched control annotations contain "stable edges" (same factors, targets retained across states) and "expression-only" sites (targets whose level changes but whose incoming connections do not).
Fourth, we partition heritability with S-LDSC on immune-mediated disease GWAS, the same class of traits Soskic et al. analyzed. Height and other non-immune traits serve as negative-control traits. The main statistic is the rewired-edge annotation's enrichment and its standardized effect size conditional on all control annotations.
Baselines
The naive reference point is the S-LDSC baseline model with no T-cell annotation at all, plus a generic "genes specifically expressed in CD4+ T cells" annotation. Any activation-related annotation must beat this.
The strongest existing competitor is the node-level approach of Soskic et al. It uses division-state-associated genes (over 2,600 genes correlated with division state) and their early-state gene programs from consensus matrix factorization as heritability annotations. This is the real threat because it already captures "early activation matters." We rebuild these annotations from the same cells and the same donors, using the same gene windows as our edge annotations. The only difference is whether a site enters because its gene's level changes or because its incoming connection changes. The file reports no enrichment numbers for this approach, so we rerun it ourselves: no reported result found on Lacuna for S-LDSC enrichment of division-state gene programs. Beating it means a significant conditional standardized effect for rewired edges with gene-program annotations included in the model.
The second competitor is the static network-hub approach of The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution, which reported that about 29% of modules were trait-specific. It gives no immune-cell enrichment value we can quote, so we rebuild local-hub annotations in two ways. One uses GTEx whole blood as in the paper. The other uses a single network pooled over all division states, which isolates the effect of time from the effect of cell resolution.
The key ablation replaces division-state grouping with grouping by physical time since stimulation, which mixes division states, and then with total pseudo-bulk, which mimics bulk tissue. It keeps everything else identical, including cell counts, which are subsampled to match. This ablation isolates the heterogeneity-masking claim. If masking matters, enrichment should shrink from division-state networks to time-point networks to bulk networks.
Why not just…?
Why not just use the division-state expression programs as annotations? Rewired edges and expression-changing genes overlap heavily, because a factor that gains a target often raises its level. Edge enrichment could therefore simply restate the Soskic result. We treat this as the main threat. The conditional S-LDSC test is designed to settle it, and so is the "expression-only" control annotation, which contains targets whose level changes with stable inputs. PANDA-based work shows factors changing targets without changing their own expression (Lopes‐Ramos et al.), so the two sets should separate enough to test. If conditional enrichment vanishes, we report that edges add nothing beyond nodes.
Why not just use static network hubs from an immune cell type? Taylor et al. showed that network structure concentrates heritability, and a pooled T-cell network might capture most of it. The pooled-network baseline tests this directly. Our claim requires that the rewired-edge annotation remain significant when static local hubs from the pooled network are in the model.
Why not just model interactions genome-wide with i-LDSC? Discovering non-additive heritability using additive GWAS summary statistics detected interaction heritability in 24 of 25 UK Biobank traits using a fixed ±50-variant window. That window says nothing about which regulatory connection the interacting variants share, which is exactly the question here. We use i-LDSC only as a secondary analysis, restricting interaction pairs to factor-motif and target-promoter variants on rewired edges versus matched stable edges. We make no primary claim from it, because such trans-acting pairs fall outside the cis window where the method's calibration was shown.
Why not just deconvolve bulk time courses instead of using single cells? Count Bridges enable Modeling and Deconvolving Transcriptomic Data outperformed CIBERSORTx and MuSiC at recovering cell proportions. However, deconvolution references lack transient states, as the POPPY authors note for their own method (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level), and division states are exactly such states. We therefore use deconvolution only for later transfer to bulk cohorts, not for discovery.
Proposed Experiments and Decisive Checks
The primary experiment runs conditional S-LDSC for the rewired-edge annotation across immune-mediated disease GWAS. The model includes the baseline, T-cell specific expression, division-state gene programs, stable edges, and pooled-network hubs. We expect significant positive conditional effects concentrated in transitions between early division states, consistent with Soskic et al.'s early-state finding. We expect none for height.
The null is a permutation that swaps edges among target genes matched on expression level, expression change and motif density. This prevents enrichment from arising merely because rewired targets are highly expressed immune genes.
The heterogeneity ablation, going from division-state to time-point to bulk networks, should show enrichment decreasing monotonically. A flat curve would falsify the masking argument.
For external replication, we use OneK1K single-cell eQTLs and assign division states to cells with CellDivider. Variants on rewired edges should show eQTL effects that vary with predicted division state more often than variants on matched stable edges.
The idea is abandoned if either of two things happens. The first is that conditional enrichment disappears once gene programs are included. The second is that rewired edges fail to replicate across donor splits at the factor level with the stability MONSTER achieved across cohorts.
References
- The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution
- Division state reveals hidden genetic regulation during T cell activation and identifies immune disease–linked gene programmes
- Estimating drivers of cell state transitions using gene regulatory network models
- Regulatory network changes between cell lines and their tissues of origin
- Discovering non-additive heritability using additive GWAS summary statistics
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level