Predicts haplotype-specific 3D genome organization and Hi-C contact maps from a single long-read Fiber-seq assay, using no DNA sequence as input.
No providers recorded yet. Browse all providers
Hi-C and Micro-C measure how the genome folds, but they average over both copies of every chromosome. Fewer than 10% of Hi-C reads can be assigned to a haplotype, so the two parental copies of a locus — which may differ in accessibility, methylation, or structure — collapse into one contact map. That matters most at structural variants, in repetitive regions, and on the X chromosome, where the active and inactive copies fold differently but have been studied almost entirely in mice.
FiberFold, from the labs of Nicolas Altemose at Stanford University and Andrew Stergachis at the University of Washington, takes a different route. Fiber-seq, a PacBio long-read assay, reports chromatin accessibility, CpG methylation, and CTCF occupancy on individual DNA molecules long enough to phase. FiberFold maps those signals — plus CTCF motif orientation — onto a Hi-C contact map at 10 kb resolution across a roughly 2 Mb window, so one sequencing run yields accessibility, methylation, protein occupancy, genetic variation, and 3D structure together.
The model deliberately excludes DNA sequence. Sequence-to-structure genome models generalize poorly across cell types; conditioning only on measured epigenomic state leaves FiberFold free to predict differently for two haplotypes or two cell types sharing a reference sequence.
The architecture adapts C.Origami: an input window of 2,097,152 bp passes through 13 one-dimensional residual convolution blocks compressing it to an 8,192 bp latent resolution, then 8 transformer blocks capturing long-range relationships, then 5 two-dimensional dilated residual convolution blocks that emit the 10 kb contact map. Hi-C targets are log-transformed and input tracks depth-normalized to a 0-1 range; training minimizes mean squared error against experimental maps. Data came from GM12878 Fiber-seq paired with published Hi-C, tiled as 2 Mb windows on a 36 kb sliding step with ENCODE blacklist regions and coverage gaps removed; augmentation applied a random shift of up to 10 kb per epoch, Gaussian noise, and a 50% chance of reversing both matrices. Chromosome 10 was held out for validation and chromosome 15 for testing, on 4 A100 GPUs with Adam at a learning rate of 0.0004 and a batch size of 8.
On held-out chr15 the model reached a median map-to-map Pearson r of 0.938 and Spearman r of 0.906, pixel-wise Pearson r of 0.932, and a median insulation correlation of 0.836 at a 500 kb radius. Transferred to K562 without retraining, map-to-map correlations held at 0.930 Pearson and 0.902 Spearman, with pixel-wise r of 0.871. Haplotype-averaged predictions matched experimental Hi-C at r = 0.938 on autosomes and 0.91 on chrX, and the per-haplotype maps showed the inactive maternal X losing sub-megabase TADs and distal contacts relative to the active paternal copy — attenuation, not the near-complete TAD loss reported in mice.
The immediate use is rare-disease genomics. For patient UDN318336, carrying a 46,XX,t(X;13)(p22.1;q14.1) translocation, phased contigs spanning both breakpoints yielded a sawtooth pattern of TAD differences between the derivative and intact X chromosomes, showing how a structural variant reorganizes local chromatin. More broadly, any laboratory already generating Fiber-seq data can add 3D structure to its readout without a separate Hi-C experiment — useful for variant prioritization, allele-specific regulation, and regions short reads cannot map.
FiberFold is a bioRxiv preprint. Its central claim — that accessibility, CTCF binding, and CpG methylation alone suffice to reconstruct large-scale chromatin structure at accuracies comparable to sequence-based models — reframes what the sequence encoder in this model class contributes, and its human X-inactivation result shows a question phased predictions can answer and bulk Hi-C cannot. Training and prediction code is on GitHub, and the pretrained checkpoint — a 172 MB file in a lab-hosted S3 bucket — is retrievable from the bucket's object endpoint, though the wget command in the README points at an AWS console URL and fails. The repository carries no license file, so reuse terms are undefined. The model is fixed to one window size and one resolution and requires Fiber-seq input, tying adoption to PacBio workflows; the authors point to Micro-C training data as the route past the first constraint.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.