bio.rodeo
ModelsOrganizationsLeaderboardAbout
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
DNA & Gene foundation models
DNA & Gene

ARCH3D

University of Michigan

Foundation model for 3D genome architecture, using masked locus modeling over genome-wide contact profiles to capture chromosome-scale organization.

Released: February 2026

ARCH3D is a foundation model for the three-dimensional architecture of the genome, developed by researchers at the University of Michigan and posted to bioRxiv in early 2026. The spatial folding of chromosomes — captured experimentally by Hi-C and related chromosome-conformation assays — encodes how distant regulatory elements and genes come into contact, but most computational models of this data operate locally or on individual contact maps. ARCH3D instead aims to represent genome structure at a global, genome-wide scale.

The model is trained with a masked locus modeling objective: it learns to predict the genome-wide contact profiles of loci that are masked out, using the surrounding profiles spread across the entirety of the genome. This self-supervised approach lets the model build representations that capture long-range, chromosome-spanning organization rather than only nearby interactions.

By preserving spatial genomic organization and reconstructing chromosome-scale interactions even under limited data, ARCH3D is positioned as a building block toward computational systems that can simulate genome structure and function.

#Key Features

  • Genome-wide scope: Models global, chromosome-spanning genome architecture rather than only local contact neighborhoods.
  • Masked locus modeling: A self-supervised objective that predicts genome-wide contact profiles of masked loci from the rest of the genome.
  • Robust reconstruction: Reported to reconstruct chromosome-spanning interactions even when training data are limited.
  • Multi-way interactions: Capable of identifying higher-order, multi-way contacts beyond pairwise interactions.

#Technical Details

ARCH3D is a foundation model trained via masked locus modeling on genome-wide contact profiles, learning to infer the contact profile of masked loci from those distributed across the whole genome. The preprint reports that the model preserves spatial genomic organization, reconstructs chromosome-spanning interactions despite limited data availability, and identifies multi-way interactions. The work is released under a CC BY license. As a recent preprint, specific architectural details, parameter counts, the training datasets used, and the availability of released weights and code should be confirmed against the manuscript.

#Applications

ARCH3D is intended for genome-biology and computational-genomics researchers studying 3D genome organization. Potential uses include representing and reconstructing chromosome-scale contact structure from sparse or incomplete measurements, detecting multi-way chromatin interactions, and providing learned embeddings of genome architecture that can support downstream analyses or simulations of genome function.

#Impact

ARCH3D extends the foundation-model paradigm from genomic sequence to the spatial, 3D organization of the genome, addressing a global modeling gap left by methods that work locally or per-contact-map. As a recent preprint, its broader adoption and independent validation remain to be established, but it points toward foundation models that could underpin simulations of how genome structure shapes function.

Citation

ARCH3D: A foundation model for global genome architecture

Galioto, N., et al. (2026) ARCH3D: A foundation model for global genome architecture. bioRxiv.

DOI: 10.64898/2026.02.23.707580

Recent citations

Papers that recently cited this model.

Not enough citation data yet.

Top citations

The most-cited papers that cite this model.

Not enough citation data yet.

Related models

Models with similar goals, methods, or subject matter.

  • Evo2HiC

    University of Washington

    Multimodal foundation model that distills Evo 2 into a compact encoder guided by Hi-C data, predicting cell-type-specific 3D genome architecture.

    DNA & GeneSpatial omics
  • ChIANet

    Central South University

    Multimodal deep learning model that predicts protein-mediated chromatin contact maps and loops de novo from protein-binding profiles and sequence.

    DNA & Gene
  • Atacformer

    University of Virginia

    Transformer foundation model for single-cell ATAC-seq that embeds both cells and cis-regulatory elements for annotation and batch correction.

    Single-cellDNA & Gene
  • D3LM

    Renmin University of China

    DNA foundation model using masked discrete diffusion to unify bidirectional sequence understanding and de novo generation in one architecture.

    DNA & Gene
  • Nucleotide Transformer v3 (NTv3)

    InstaDeep / Research Institute of Molecular Pathology (IMP) / Medical University of Vienna / Cornell University / Cold Spring Harbor Laboratory

    Multi-species genomics foundation model spanning representation learning, functional-track prediction, and sequence generation at 1 Mb context.

    DNA & Gene

Citations

Total Citations0
Influential0
References0

Fields of citing research

Not enough data

Openness

bio.rodeo opennessClosed · low usability and reproducibility
19Closed
Usability — can I run it?13
Reproducibility — can I retrain it?13
Model Openness Framework
Unclassified
Missing required components

Tags

3d_genomechromatincontact_map_reconstructionfoundation_modelgenome_architecture_modelingrepresentation_learningself_supervisedtransformer

Resources

Research Paper