Model Architecture
We developed Origami, a general model architecture based on CNN and transformers, to empower genomics studies. Origami incorporates both nucleotide-level DNA sequence and genomic signals (such as ATAC-seq signal) to make accurate, cell-type-specific predictions of genomic features. In this study, we demonstrate the model variant for predicting 3D chromatin architecture – C.Origami.
Cell type-specific Chromatin Architecture Prediction
Given the DNA sequence, CTCF ChIP-seq and ATAC-seq signal, C.Origami makes de novo prediction of the Hi-C contact matrix within a 2Mb window. The prediction can be completed under 1 second with GPU.
In silico Genetic Perturbation and Screening
The high accuracy of prediction enables systematic in silico genetic studies of the 3D genome for cell type-specific factors – both cis-regulatory elements and trans-acting factors – that contribute to chromatin structures. Please refer to the full text for more applications and in silico experiments enabled by C.Origami.
Live Demo
Interactive live demo for IMR-90 and GM12878 prediction in real time. Reference genome fasta, cell type-specific CTCF ChIP-seq and ATAC-seq bigwigs are used as inputs.
Notes for the live demo:
- The app is in alpha phase and might be unstable.
- The next version will support custom sequence and bigwig/peaks inputs. Stay tuned!
Full Text
Code Availability
Code is currently beening documented and will be released shortly.