Loading…
Infers gene-centered chromatin interactions from bulk RNA-seq alone, mapping 3D genome changes across 12,347 tumor and normal transcriptomes.
No providers recorded yet. Browse all providers
A Hi-C library deep enough to call individual loops costs far more than an RNA-seq run, so the 3D genome is absent from nearly every large clinical cohort that already has a transcriptome: TCGA holds expression profiles for more than ten thousand tumors and matched conformation data for essentially none of them. AI4Loop attacks that gap from the cheap side, taking the expression matrix a cohort already has and predicting which pairs of gene promoters are in spatial contact.
The target is what the authors call a gene-centered chromatin interaction (GCI) — an intrachromosomal loop whose two anchors both overlap annotated gene bodies, coupling promoters to each other rather than tying an enhancer to a single promoter. The premise is that genes sharing a nuclear hub share coordinated transcriptional behavior, so the expression landscape surrounding two loci carries information about whether they touch. AI4Loop reads that landscape at seven spatial resolutions at once and emits an interaction probability per candidate gene pair.
The framework comes from Melissa Fullwood's group at Nanyang Technological University, posted as a preprint in August 2026. It succeeds the same lab's ChINN, which called loops from DNA sequence. The swapped modality is the point: sequence-based predictors describe a reference genome's folding potential, while an RNA-seq-driven model tracks state that differs between one tumor and the next. That also separates AI4Loop from cataloged neighbors needing assays most cohorts lack — Polaris and TRUHiC take contact maps, UniversalEPI sequence plus ATAC-seq.
AI4Loop is a dual-branch bidirectional LSTM: each anchor's multi-scale expression vector runs through its own Bi-LSTM encoder, and the two encodings are concatenated and passed to fully connected layers with a sigmoid output. Training labels come from HiCCUPS loop calls on the Rao et al. Hi-C datasets for four ENCODE cell lines, restricted to intrachromosomal gene-gene pairs 5 kb to 2 Mb apart, with negatives from four complementary sources at a 1:5 ratio. Random-split evaluation reaches AUPRC above 0.5871 and AUROC above 0.8199; holding out chromosomes 4, 7, 8 and 11 gives AUPRC above 0.5363 and AUROC above 0.8467, against a 0.17 prevalence baseline. Training on one cell line and testing on another holds AUPRC above 0.47, and on loops unique to the target cell type AUPRC runs 0.53–0.73. The architecture beat logistic regression, a feedforward network, a CNN, an RNN, a plain LSTM and a transformer on AUPRC in every cell line.
Accuracy depends strongly on gene count: genome-wide input exceeding 60,000 genes yields AUPRC above 0.55, while restricting to 1,000 genes drops it to roughly 0.23. On six chronic lymphocytic leukemia patients with matched Hi-C, the GM12878 checkpoint scored 61–67% and the K562 checkpoint exceeded 88%. Only the random-split checkpoints are distributed, the more optimistic of the paper's two protocols.
The intended use is retrofitting 3D genome analysis onto transcriptomic resources. Across 12,347 TCGA samples from 32 cancer types, AI4Loop scored 49,255 candidate pairs among 5,338 robustly expressed genes and found elevated GCI counts in 97.38% of tumors, concentrated at oncogene loci. A K562-driven pass over 3.59 million pairs in TARGET-AML separated leukemia from healthy marrow at ARI 0.8988. Screening 57,353 CLUE L1000 drug-response profiles across 42 cell lines nominated compounds predicted to reverse those gains, and Hi-C on treated MCF7 cells confirmed that the oxazolidinone antibiotics eperezolid and radezolid do reduce them.
AI4Loop's contribution is the demonstration that coordinated expression carries enough structure to recover chromatin contacts at cohort scale, turning a pan-cancer 3D genome survey into a computation over data that already exists. The scope is deliberately narrow — intrachromosomal gene-gene contacts between 5 kb and 2 Mb, with inter-chromosomal and enhancer-associated interactions named as future work and single-cell use limited to pseudo-bulk aggregation. Code and checkpoints carry an NTUitive dual license permitting non-commercial use only, and the work remains a preprint awaiting peer review.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.