Loading…
Chromatin interaction prediction from DNA sequence alone, calling CTCF-, RNA Pol II- and Hi-C-associated loops between open chromatin regions.
No providers recorded yet. Browse all providers
Calling chromatin loops genome-wide by experiment is expensive in cells as much as in reads: Hi-C and ChIA-PET libraries deep enough to resolve individual interactions need material that a frozen vial of leukemia cells from a patient biopsy rarely provides. The 3D genome is consequently well mapped in a few cultured cell lines and barely mapped across clinical cohorts. ChINN attacks that by predicting which pairs of open chromatin regions interact using only the DNA sequence beneath the two anchors and the distance between them, so that one accessibility assay plus a reference genome is enough to annotate loops across an entire sample.
ChINN was developed by Fan Cao, Yu Zhang, Yichao Cai and colleagues in Melissa Fullwood's group at the Cancer Science Institute of Singapore, with collaborators at Nanyang Technological University, A*STAR, the National University of Singapore and the National University Cancer Institute, Singapore. It was posted to bioRxiv in July 2019 as a CTCF and RNA Pol II model, extended with Hi-C models in a December 2020 preprint, and published in Genome Biology in August 2021.
Earlier sequence-based predictors were mostly confined to enhancer-promoter pairs, fixed distance windows, or reconstructing Hi-C contact matrices. ChINN instead emits discrete loop calls at unrestricted anchor separation — the output a differential-interaction analysis consumes, and a different product from contact-map predictors such as Polaris and ChIANet.
Each anchor passes through a three-layer 1D convolutional feature extractor with width-8 kernels, max pooling and a weighted-sum pooling step that yields a fixed-length embedding. Training runs in three stages: the convolutional extractor is fitted first on distance-matched datasets (roughly 1:5 positive-to-negative) with fully connected layers as the classifier, then frozen while gradient-boosted trees are trained on extended negative sets and finally on all anchor pairs generated from open chromatin. Anchor distance is an optional classifier feature: uninformative alone, it adds signal in combination with the sequence features. Within-sample auPRC reaches 0.77 for GM12878 CTCF and 0.75 for HelaS3 CTCF, 0.47–0.60 for the three RNA Pol II models, and 0.52–0.77 across the eight Hi-C cell lines. The harder from-open-chromatin task scores 0.514 (CTCF) and 0.347 (RNA Pol II) against positive-to-negative ratios near 1:122 and 1:186. Applied unchanged to six newly generated CLL patient samples, the GM12878 Hi-C model reaches auPRC 0.2772–0.4362 and the K562 model 0.2607–0.3996, with decision thresholds of 0.025 and 0.016 chosen by F-score. Models trained on CTCF interactions transfer poorly to RNA Pol II interactions and vice versa, which the authors attribute to their different sequence contexts. The implementation is PyTorch, released under Apache 2.0 with the ENCODE-derived training data and the preprocessing, training and prediction scripts in the repository.
ChINN suits cohort-scale 3D genome annotation where conformation capture is impractical: given ATAC-seq or DNase-seq peaks and the reference genome, a fixed checkpoint annotates loops for every sample in a study. Applied to 84 previously published CLL ATAC-seq samples, it predicted 48,443 CTCF-associated, 23,633 RNA Pol II-associated and 758,407 Hi-C-associated open chromatin interactions, exposing widespread patient-specific looping and systematic differences between IGHV-mutated and IGHV-unmutated disease at prognostic genes including LPL and CD38. Predictions in MCF-7 cells were confirmed by 4C-seq at the GREB1 and SIAH2 loci, including interactions the matched ChIA-PET library had missed.
ChINN established that chromatin interactions between accessible regions can be predicted from primary sequence at genome scale, and that models fitted on cell lines carry over to clinical material — which is what makes interaction heterogeneity measurable across a patient cohort rather than one locus at a time. The authors are candid about the cost: cross-sample auPRC of roughly 0.26–0.6 leaves substantial error, inevitable given how far negatives outnumber positives genome-wide, and predictions are confined to pairs of open chromatin regions, so an accessibility assay is still required.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.