Orca

A 40.5-Mb inversion on chromosome 3 moves a GATA2 enhancer next to the EVI1 promoter and causes acute myeloid leukemia. Reading that off the sequence needs an input window larger than the rearrangement, and the sequence models of chromatin folding that preceded Orca — Akita and DeepC — had topped out at 1-Mb distance, too small to represent such an inversion at all. Orca reads up to 256 Mb of DNA, more than chromosome 1's 249 Mb, and returns nested 250 × 250 contact matrices at nine scales.

The mechanism is a hierarchical encoder feeding a cascade of decoders. The encoder passes bottom-up over one-hot sequence, coarsening it to 4, 8, 16 … 1,024 kb, then top-down to return long-range context to the finer levels. Each decoder predicts one window size from its own representation, a distance-based background matrix, and the level above it, upsampled and cropped — so the output is literally a zoom, the 1-Mb map refined out of the 2-Mb map. Scores are log fold changes over the average contact at that distance, separating structure from Hi-C's distance decay.

Orca is the work of Jian Zhou, sole author, in the Lyda Hill Department of Bioinformatics at UT Southwestern Medical Center, published in Nature Genetics in May 2022. The encoder's first section is adapted from Sei, the same author's chromatin-profile model and a descendant of Beluga, carrying a dual linear-plus-nonlinear residual design built for regulatory tracks over to contact maps.

#Key Features

• Nine scales from one input: Predictions come at 1, 2, 4, 8, 16, 32, 64, 128 and 256 Mb, and interchromosomally at the 32–256-Mb levels when the input spans several chromosomes.
• Variants of nearly any size: Deletions, duplications, inversions, insertions, translocations and arbitrary multi-segment rearrangements are scored by rebuilding the mutant chromosome and re-predicting. Tested variants ran from 300 bp to 80 Mb, seconds apiece.
• Horizontal checkpointing: Bottom layers run in blocks sized to their receptive field — 212 kb for the 4-kb output, under 1/1,200 of a 256-Mb input — so peak memory stops scaling with sequence length, which is what puts chromosome-scale training within reach of four GPUs.
• Virtual genetic screens: Multiplexed in silico mutagenesis disrupts many 10-bp sites at once and deconvolves per-site effects across three random disruption sets, a 20-fold speedup at above 0.9 correlation with one-at-a-time mutagenesis. The three modules load independently, so a screen runs on the cheapest one covering its range.

#Technical Details

Both halves are convolutional with residual connections. The encoder's first section stacks 28 convolution layers at 64–128 channels to reach 4-kb resolution, then four layers per octave at 128 channels up to 1,024 kb; each decoder is a 2D residual stack of 112 dilated convolution layers fed by pairwise sums of the 1D representation. Separate models were trained on micro-C for H1-ESC (4DN accession 4DNFI9GMP2J8) and HFF (4DNFI643OYP9) against GRCh38, holding out chromosome 8 for validation and 9 and 10 for test, under mean-squared-error loss plus an auxiliary head predicting DNase-seq, CTCF and histone-mark peaks at 4 kb. Training ran in three stages of 480,000, 150,000 and 20,000 steps on four NVIDIA V100 GPUs, freezing the earlier stage's encoder each time. On held-out chromosomes the models reach 0.78–0.85 Pearson correlation with observed micro-C for H1-ESC and 0.73–0.79 for HFF across all scales, and 0.47–0.74 interchromosomally. A third model trained on cohesin-depleted HCT116 Hi-C, where TADs vanish but compartments persist, isolates compartment logic from loop extrusion.

#Applications

The immediate use is interpreting noncoding structural variants that disrupt regulatory topology. Across 14 transposon-mediated TAD-boundary insertions measured by in situ Hi-C, predicted insulation changes matched observation at cosine similarity 0.89 for H1-ESC and 0.76 for HFF, reproducing all three reported outcome classes. In KCNJ2–SOX9 it separates the three phenotypes produced by nested duplications — sex reversal, Cooks syndrome, and no phenotype — by showing which copies gain a new insulating boundary, and it explains brachydactyly, F-syndrome and polydactyly variants as three routes to one ectopic enhancer contact. Clinical geneticists gain a way to triage large rearrangements that carry no coding consequence.

#Impact

Used as an "in silico genome observatory", Orca yielded hypotheses, not just predictions. A genome-wide 10-bp screen found over 88.9% of the highest-impact sites overlap CTCF motifs, while the mid-impact band is dominated by cell-type-specific factors: POU5F1::SOX2 is 48.7-fold enriched in H1-ESC, FOSL1::JUND 167-fold in HFF. Inserting 800 bp of transcription-start-site sequence flips a B compartment to A, while 6–12 kb is needed for the reverse — the same length scale experiments report as the minimum that sustains compartments. Evo2HiC benchmarks against Orca as the reference sequence-to-Hi-C baseline. Three limits bound the release: coverage is two cell lines plus the cohesin-depleted model, so new cell types require retraining; repetitive regions cannot be validated against Hi-C; and while the Zenodo weights carry CC BY 4.0, the code is under a UT Southwestern licence limited to academic research, with commercial use expressly prohibited.