ORIGAMI

Transcription of a single gene is shaped by candidate cis-regulatory elements (cCREs) scattered hundreds of kilobases to megabases away, and those elements interact with one another rather than acting independently. Existing assays each capture half the picture: Hi-C reports physical proximity but not regulatory function, while CRISPR screens and MPRAs report function but cannot scale to every element combination in every cell type. ORIGAMI — multiOmics-based cis-Regulatory Interaction and Gene Activity Modeling with hierarchical Inference — closes that gap by casting cis-regulatory modeling as a latent graph inference problem.

Developed at Zhejiang University with collaborators at Hangzhou City University, Shanghai Jiao Tong University, City University of Hong Kong, and Iowa State University, ORIGAMI is gene-centric rather than window-centric. For each protein-coding gene it assembles a graph over the 200 nearest active cCREs — 100 upstream, 100 downstream — plus the gene's own promoter, initializes the edges from sparse and noisy Hi-C contacts, and learns a denoised interaction matrix over them. That reconstruction is coupled to a gene expression regression head trained on the same graph, so only edges that improve transcriptional prediction survive.

This design distinguishes ORIGAMI from sequence-to-function models such as Enformer and Borzoi, which read a contiguous DNA window and expose element-gene relationships only indirectly through attention or gradient attribution. ORIGAMI emits the regulatory graph itself, including indirect paths in which a distal element reaches the promoter through an intermediate element.

#Key Features

• Function-constrained graph inference: A graph autoencoder denoises Hi-C priors into a cis-regulatory interaction matrix, while a joint expression-prediction loss filters out contacts that reflect structural proximity without regulatory consequence.
• Element-centric inputs: Modeling only ENCODE-annotated active cCREs, rather than whole genomic windows, removes inactive-sequence redundancy and lets the receptive field extend across megabases — far enough to recover the seven experimentally validated MYC enhancers in K562, one of which sits 2 Mb downstream of the TSS.
• Zero-shot cross-cell-type transfer: Reconstructs regulatory interactions in cell types never seen during training without any additional fitting.
• In silico perturbation: Predicts transcriptional consequences of both cis-acting sequence variants and trans-acting transcription factor knockouts, and reports the accompanying rearrangement of the regulatory graph.
• Interpretable topology: The 11,596 graphs inferred for K562 sort into simple, sparse, and complex regimes whose modularity tracks cell identity, disease genes, and GWAS SNP enrichment.

#Technical Details

Each of the 201 regulatory elements is represented as a 350 bp window with 17 channels: one-hot nucleotide sequence, read-depth-normalized DNase-seq, and 15 histone and transcription factor ChIP-seq tracks. A hierarchical transformer encodes these in two stages — a CRE-centric encoder of two position-aware blocks with four attention heads and bucketed relative position bias, then a regulatory-centric encoder whose position bias sums learnable terms at 150, 600, 2,400, 10,000, and 40,000 bp scales, with an attention mask enforcing information flow from cCREs to the promoter. A GATv2 encoder and a Euclidean-distance decoder reconstruct edge probabilities (thresholded at 0.7), and DiffPool over GraphSAGE layers aggregates the pruned graph into a gene-level embedding regressed to expression through a softplus head. Embeddings are 256-dimensional; training minimizes a weighted sum of binary cross-entropy on edges and Poisson negative log-likelihood on log-transformed TPM.

Training used nine ENCODE cell lines with GENCODE release 30 annotations, SCREEN Registry V3 cCREs, and intact Hi-C binned to 350 bp. Chromosome 16 was held out for validation and chromosomes 8 and 9 for in-cell-type testing; all K562 autosomes were withheld entirely, making K562 chr8 and chr9 a strictly zero-shot set on which ORIGAMI reached 0.84 auROC and 0.82 auPRC for interaction reconstruction and Pearson r = 0.84 for expression. It led across expression metrics against EPCOT-CNN, EPCOT-LSTM, CREaTor, and GraphReg retrained on matched data, and outperformed ABC contact, Enformer, GraphReg, CREaTor, raw cCRE contacts, and inverse distance at prioritizing CRISPR-validated K562 cCRE-gene pairs.

#Applications

ORIGAMI suits researchers interpreting noncoding variation and regulatory architecture. Applied to fine-mapped breast mammary eQTLs it recovered effect directionality and correlated with GTEx effect sizes (Spearman rho = 0.56 near the TSS), tracing one variant's effect to a disrupted BATF3 motif. Simulated CTCF knockout reproduced the direction of change for 74% of down-regulated and 64% of up-regulated differentially expressed genes, with Pearson r = 0.45 on log2 fold changes. Contrasting networks across mammary-epithelial-to-MCF-7 transformation identified genes gaining (10.2%) or losing (14.6%) modular organization during oncogenesis, nominating hub elements for follow-up perturbation.

#Impact

ORIGAMI argues that explicit, function-constrained network reconstruction recovers regulatory logic that pairwise enhancer-gene scoring misses — notably synergistic effects, where knocking out enhancers from different modules of the MYC network collapses expression more than the sum of single knockouts. It remains a preprint awaiting peer review. Code and tutorials are released under GPL-3.0, though no trained checkpoint is distributed, so reproducing results requires retraining on the ENCODE inputs. The authors note that training on population-averaged bulk data prevents modeling continuous cell-state variation, and that gene-centric graphs capture neither feedback regulation nor trans-regulatory coupling.