Lab Newsletter — September 14, 2026: Folding the Genome

This week the digest has read the genome — designing the mRNA message, tracing the conversations between cells. But the genome is not a one-dimensional string of letters. Inside the nucleus it is packed, looped, and folded into a precise three-dimensional architecture, and that shape is not decoration: where a gene sits in the fold — which enhancers loop around to touch it, which walls of a domain contain it — is as decisive as the letters themselves. For years that architecture could only be measured, expensively, with assays like Hi-C. Today’s digest is about the deep-learning models that learned to predict how the genome folds, directly from its sequence — and what that unlocks.

🧬 The founding move: fold from sequence alone

The DNA sequence is one-dimensional; its fold is not. Fudenberg, Kelley & Pollard (Nature Methods, 2020) set out from a striking fact — “the human genome sequence folds in three dimensions into a rich variety of locus-specific contact patterns” — and the open question that “how a given DNA sequence encodes a particular locus-specific folding pattern remains unknown.” Their answer was Akita: “a convolutional neural network … that accurately predicts genome folding from DNA sequence alone.” The model didn’t just fit the data; it revealed the rules. “Representations learned by Akita underscore the importance of an orientation-specific grammar for CTCF binding sites” — the insulator protein whose direction, not just presence, sets where loops anchor. And a model you can run in silico becomes an instrument: Akita can “perform in silico saturation mutagenesis, interpret eQTLs, make predictions for structural variants and probe species-specific genome folding.” The goal, in their words, is “decoding genome function from sequence through structure.”

📏 Megabase context, and reading the variant

A fold is set by more than a single site — it takes megabases of context, and the payoff is interpreting the noncoding genome. Schwessinger et al. (Nature Methods, 2020), publishing alongside Akita, framed exactly that stake: “predicting the impact of noncoding genetic variation requires interpreting it in the context of three-dimensional genome architecture.” Their DeepC is “a transfer-learning-based deep neural network that accurately predicts genome folding from megabase-scale DNA sequence.” Transfer learning lets it see far enough to matter, and the result is both structural and clinical: DeepC “predicts domain boundaries at high resolution, learns the sequence determinants of genome folding and predicts the impact of both large-scale structural and single base-pair variations.” From a single changed base to a whole rearranged chromosome, the fold becomes the missing context that tells you what a noncoding variant actually does.

🔭 Multiscale: kilobase to whole chromosome

The fold has structure at every scale — loops of a few kilobases, domains, chromosome-spanning compartments — and a real model should span them all. Zhou (Nature Genetics, 2022) delivered it with Orca, “a sequence-based deep-learning approach … that predicts directly from sequence the 3D genome architecture from kilobase to whole-chromosome scale.” It captures the full vocabulary: “chromatin compartments and topologically associating domains, as well as diverse types of interactions from CTCF-mediated to enhancer-promoter interactions and Polycomb-mediated interactions with cell-type specificity.” And it reads variants across five orders of magnitude — Orca “recapitulated effects of experimentally studied variants at varying sizes (300 bp to 90 Mb)” — while enabling “in silico virtual screens to probe the sequence basis of 3D genome organization at different scales.” One model, the whole hierarchy of the fold.

🎯 Cell-type-specific — and screenable in silico

The same genome folds differently in a neuron and a T cell, and measuring every cell type by Hi-C is prohibitive. Tan et al. (Nature Biotechnology, 2023) named the bottleneck — “experimental methods for measuring three-dimensional chromatin organization, such as Hi-C, are costly and have technical limitations, restricting their broad application particularly in high-throughput genetic perturbations” — and answered with C.Origami, “a multimodal deep neural network that performs de novo prediction of cell-type-specific chromatin organization using DNA sequence and two cell-type-specific genomic features—CTCF binding and chromatin accessibility.” On top of prediction they built discovery: “an in silico genetic screening approach to assess how individual DNA elements may contribute to chromatin organization and to identify putative cell-type-specific trans-acting regulators.” Applied “to leukemia cells and normal T cells,” it shows such screens “can be used to systematically discover novel chromatin regulation circuits in both normal and disease-related biological systems.” Cheap, high-throughput hypotheses where the wet assay can’t go.

🧪 Generalize across cell types, from cheap tracks

There’s a catch the field kept hitting. Yang et al. (Genome Biology, 2023) put it plainly: “recent deep learning models that predict the Hi-C contact map from DNA sequence achieve promising accuracy but cannot generalize to new cell types.” Their Epiphany shifts the input to what labs already have — “a neural network to predict cell-type-specific Hi-C contact maps from widely available epigenomic tracks.” It “uses bidirectional long short-term memory layers to capture long-range dependencies and optionally a generative adversarial network architecture to encourage contact map realism,” and, crucially, it travels: Epiphany “shows excellent generalization to held-out chromosomes within and across cell types, yields accurate TAD and interaction calls, and predicts structural changes caused by perturbations of epigenomic signals.” From one bespoke model per genome toward portable, cell-type-aware prediction.

🕸️ Closing the loop: structure → expression, via graph nets

The fold matters because it decides which enhancers reach which genes — so the real prize is turning 3D contacts back into gene expression. Karbalayghareh, Sahin & Leslie (Genome Research, 2022) took aim at exactly that “longstanding unresolved” problem of “linking distal enhancers to genes and modeling their impact on target gene expression.” Their GraphReg “exploits 3D interactions from chromosome conformation capture assays to predict gene expression from 1D epigenomic data or genomic DNA sequence.” The architecture fits the biology: “by using graph attention networks to exploit the connectivity of distal elements up to 2 Mb away in the genome, GraphReg more faithfully models gene regulation and more accurately predicts gene expression levels than the state-of-the-art deep learning methods for this task.” And it stays falsifiable — “feature attribution used with GraphReg accurately identifies functional enhancers of genes, as validated by CRISPRi-FlowFISH and TAP-seq assays,” outperforming both CNNs and the activity-by-contact model. The fold isn’t the end point; it’s the wiring diagram for expression.

🧫 Why it’s our kind of problem

Read across the six and the arc is one the lab keeps betting on: from single-purpose CNNs toward multiscale, cell-type-specific, graph-based models that connect structure to function. Akita and DeepC prove the fold is written in the sequence; Orca spans every scale; C.Origami makes it cell-type-specific and screenable; Epiphany makes it portable; GraphReg turns the 3D contact graph into a graph attention network that predicts expression. That last turn — biology becoming a graph, and representation learning turning out to be the right instrument — is the same movement we saw in the knowledge-graph digest and in yesterday’s cell-cell-communication GNNs. The in-silico genetic screens (Orca, C.Origami) are exactly the cheap, hypothesis-generating dry-lab experiments that pair with the lab’s self-driving-lab ambitions, and these models are open and measured on shared Hi-C/Micro-C benchmarks — the publish-the-model-and-the-data spirit behind the BioImage Model Zoo and BioEngine. Above all it’s a load-bearing layer for the virtual cell: a faithful cell model has to know not just its gene sequences but how the genome is packed and looped in each cell type, and how a single variant reshapes that architecture. Sequence → structure → function, end to end — and, increasingly, predictable.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits; a hypha-search surrogate sweep surfaced nothing breaking. Anchors were verified via NCBI E-utilities.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.