New Biotech Model Predicts 3D Genome Folding
Researchers introduced HiFold-Mamba to improve how AI predicts complex DNA structural maps from genetic sequences.
Updated on Oct. 30, 2026 in Biotech
Scientists have developed a new reverse-complement-equivariant state-space model, HiFold-Mamba, to enhance the prediction of 3D genome folding. The study aims to determine if incorporating map-level insulation constraints improves the accuracy of sequence-based contact maps.
Why it matters
Current sequence-to-contact-map models rely on pixel-wise losses that may overlook critical domain features. This research assesses whether insulation-based loss functions provide a more accurate alternative for predicting genomic structures against independent annotations.
HiFold-Mamba predicts 448 x 448 maps at 2,048 bp resolution from 1,048,576 bp of DNA input. It utilizes a combination of a convolutional tower, bidirectional Mamba blocks, a multi-scale mixer, and a symmetric pairwise head architecture.
The players
HiFold-Mamba
This is a new reverse-complement-equivariant state-space model designed to process DNA sequence data for genome folding predictions.
DNALongBench
This is a specialized task platform used by researchers to benchmark and test the performance of various computational models in genomics.
The details
The model was tested using a paired bootstrap method across 237 one-megabase blocks to evaluate its performance against three trained baselines. Observed maps from validation windows successfully recovered independent boundaries with an AUROC between 0.61 and 0.91, while loop recovery reached an average precision of 0.86 to 0.90.
Timeline
The article detailing the model was published on bioRxiv on October 30, 2026.
The Big Picture
This study challenges the dominance of pixel-wise loss functions in genomic modeling by testing the efficacy of map-level insulation constraints. It marks a shift toward integrating structured biological knowledge directly into the architectural design of deep learning models for DNA folding.
For researchers and software developers, this model offers a new architectural approach to processing massive DNA input sequences. Its development may eventually lead to higher resolution tools for analyzing genetic disorders and understanding the fundamental mechanics of genome structure.
The takeaway
Advancing genome folding predictions requires moving beyond simple pixel-wise training toward models that account for biological structures like boundaries and loops. Future progress in this field will likely hinge on whether few-shot adaptation can consistently outperform training models from scratch.
Further reading
Explore the latest developments in Biotech to understand how computational models are shaping the future of genomic research.
More information
Access the full text of registered report on the bioRxiv portal.
