Integrating Single-Molecule Sequencing and Deep Learning to Predict Haplotype-Specific 3D Chromatin Organization in a Mendelian Condition

The diploid human genome is regulated by the complex interplay of CpG methylation, protein-DNA interactions, chromatin accessibility, and 3D genome folding along each of the two non-identical haplotypes. Existing short-read sequencing-based methods can quantify these epigenetic features individually, but they struggle to resolve individual haplotypes or structurally variable regions associated with disease. Building on recent advances in genomic deep learning and long-read sequencing, we present FiberFold, a computational method for predicting 3D genome organization while simultaneously measuring genetic variation, CpG methylation, protein-DNA interactions, and chromatin accessibility. Using data from a single multi-omic long-read sequencing experiment, Fiber-seq, we demonstrate that FiberFold can accurately predict 3D genome structure in a de novo, cell-type-specific, and haplotype-specific manner. Applying FiberFold to Fiber-seq data from a 46,XX cell line with allelic X-inactivation, we show that TADs are attenuated on the inactive chrX. Furthermore, by applying FiberFold to a Mendelian condition caused by a 13;X balanced translocation, we predict significant changes in the 3D structure of homologous sequences involved in this translocation. FiberFold showcases the power of integrating long-read sequencing with deep learning tools, enhancing its utility for investigating fundamental chromatin biology as well as the molecular basis of rare diseases.