From enhancer networks to base-resolution variant scoring

I picked these three together because they come at the same problem from three directions. The question underneath all of them is the one my PhD lives on: of the thousands of non-coding variants we can point at, which ones actually matter, and how? Paper one is functional: it finds the enhancers and the genes they control. Paper two is structural: it asks, at a single base, whether a variant changes transcription-factor binding. Paper three is predictive: it scales a chromatin model across 132 cell types and adds evolutionary constraint to prioritise variants genome-wide. Find the elements, score a variant, then prioritise across the whole genome.

Paper 1, the enhancer network (Wang & Hon, 2025)

GWAS and whole-genome sequencing have linked thousands of non-coding variants and enhancers to breast cancer, but almost none have been functionally characterised, so we don’t know how most of them contribute to the disease. Two things make it hard. Assigning an enhancer to its target gene isn’t obvious. It’s often not the nearest gene. And enhancers can act indirectly: an enhancer changes a transcription factor, and it’s that factor, acting on a distant gene, that produces the effect.

So Wang and Hon do the functional experiment at scale. They took 3,513 regulatory elements (2,123 putative enhancers, 1,390 promoters) associated with breast cancer, chosen from PCAWG somatic mutations (kept if mutated in ≥2 of 123 patients) and a GWAS meta-analysis (4,452 variants across 142 loci), filtered to open chromatin with an H3K27ac signal. They switched each element off with dCas9-KRAB CRISPRi in an ER-positive and an ER-negative line, and read the transcriptome in over 500,000 single cells (Perturb-seq), reaching ~80% repression and ~1,000 cells per element.

To assign targets they wrote pySpade, which compares perturbed against unperturbed cells with a random-sampling background correction (guides aren’t evenly distributed across cell states, which would otherwise manufacture false positives). In the ER-positive cells, 147 elements had a direct target within 2 Mb (including the cancer genes BRCA2 and MUC1) and the somatic variants were enriched in MTF1 and TP53 motifs. One clean example: a patient somatic variant in the CITED4 enhancer disrupts a putative MYC binding site.

The network, and the motif that stuck with me

Including indirect targets, the numbers explode: 25,287 enhancer-gene pairs forming a genome-wide regulatory network, spanning all chromosomes, with the median enhancer wired to about nine genes. Within it they describe two recurring shapes: one-to-multiple (one enhancer, several genes in a pathway, e.g. through PDS5B into the DREAM cell-cycle programme) and multiple-to-one, where several enhancers converge on the same gene.

Their headline example is TP53: three enhancers on different chromosomes each reach it indirectly, and they validated two by qPCR: a chromosome 17 enhancer acting through MED1, and a chromosome 6 enhancer acting through SERAC1. Finally, they repeated the whole screen in the ER-negative line and found ~72% of enhancers are subtype-specific, with distinct networks.

Paper 2, structure scores the variant (Gerasimavicius, Biddie & Marsh, 2026)

Non-coding SNVs that change transcription-factor binding can shift gene expression and cause disease. We already have good sequence-based predictors (position weight matrices, deltaSVM, DeepSEA), but they depend on training data and carry TF-specific biases. Structure could add a mechanism rather than just a score, but until recently there were almost no TF–DNA structures to work from. AlphaFold 3 changes that: it can model protein–DNA complexes.

So the question is whether AF3 structures can recapitulate measured allele binding preferences. They built 75,134 AF3 complex models (reference and alternative allele for every SNP pair) across six transcription factors (SPIB, ETV4, ELK3, SF-1, PAX5, MEIS2), benchmarked against SNP-SELEX, an in-vitro assay that labels each SNP pair as preferentially bound or not. Each model is a single TF docked onto a 40-bp DNA duplex, scored five ways: two FoldX physics energies, a DNA solvent-accessibility term, and the one that matters most: ΔipTM.

All three main metrics separate the SNP-SELEX groups, and ΔipTM gives the cleanest separation: AUC 0.685 for detection (is there any binding difference?) and 0.792 for direction (which allele wins?). So it’s better at calling the winner than at calling that a difference exists at all. Strikingly, ΔipTM, really a by-product of the model’s confidence, beats the physics-based FoldX energies, and it stays informative even when models fall below the usual 0.6 ipTM threshold.

The honest limits: it works for some TFs, not others. Good for the ETS family (SPIB, ELK3) and SF-1, and a consensus of metrics recapitulated several clinical variants (two NKX2-5 variants, a GATA3 variant at the neuroblastoma LMO1 locus, a USF2A variant). It breaks for PAX5 and MEIS2, which bind cooperatively or through multiple domains but were modelled as monomers, and it called a PAX6 enhancer variant in the wrong direction. Where good training data exist, deltaSVM is still more accurate. The authors’ own framing: a proof of concept (training-free, mechanistic, visual), but post-hoc structural support, not a genome-wide predictor. And across the board, a consensus of metrics beats any single one.

Paper 3, prioritising at genome scale (Marderstein & Montgomery, 2026)

Disease risk keeps mapping to the non-coding genome, but a variant’s effect depends entirely on which cell type and which developmental stage you look at, and we have no regulatory code to read non-coding variation the way we read the genetic code for proteins. Marderstein and Montgomery ask whether a deep-learning chromatin model can predict variant effects across all those contexts, and whether evolutionary constraint can tell us which of them matter.

They took ChromBPNet (a convolutional network that predicts chromatin accessibility base by base from DNA sequence) and trained it across single-cell ATAC-seq from 132 fetal and adult brain and heart cell types, then predicted the effects of 22 million variants per cell type. Then the prioritiser: FLARE combines those predicted effects with base-level evolutionary constraint (phyloP), flagging variants with an unusually large regulatory effect for how constrained their site is.

The findings hang together nicely. Large-effect variants in the fetal brain tend to be rare and constrained: evolution has been selecting against them. Rare variants act across many cell types; common ones are more cell-type-specific. Run on de novo mutations, FLARE’s top calls were enriched in disease cases: 14 of the 16 top-scoring mutations were in cases, not controls (autism, congenital heart disease), and the common variants it prioritised were enriched for schizophrenia heritability. So the same framework works across rare and common disease genetics. The honest limits: these are predictions that still need experimental validation, the models are brain and heart only, and they don’t yet capture splicing or transcript stability.

How the three fit together

Wang and Hon give the elements: which enhancers control which cancer genes, directly and indirectly, resolved by subtype. Gerasimavicius gives a base-resolution reading of whether a single variant in one of those elements changes transcription-factor binding, with a mechanism you can look at. Marderstein and Montgomery give the prioritisation engine: a chromatin deep-learning model plus evolutionary constraint that scales across cell types and separates the variants that matter from the ones that don’t. Elements, mechanism, prioritisation: every one of them lands on a different objective of the breast-specific, non-coding driver map I’m building.