Co-folding with a Soup of Representations

Abstract

Co-folding models such as AlphaFold3, Protenix, ESMFold2, and OpenDDE have advanced rapidly, yet no single model consistently performs best across all biomolecular complexes. In this paper, we show that their pair representations encode complementary information that can be transferred across models to improve structure prediction. We introduce SoupFold, which combines pair representations from multiple co-folding models in a common representation space and generates structures from the combined representation. Importantly, SoupFold does not retrain the co-folding models and learns only simple mappings to transfer representations across models. We evaluate SoupFold on antibody-antigen, protein-protein, protein-ligand, molecular glue, GPCR, and oligomeric complex prediction using AlphaFold3, Protenix, ESMFold2, and OpenDDE. By combining representations across models, SoupFold improves over individual co-folding models across the considered benchmarks.

Detailed Insight into the Chignolin Folding Process from Maximally Informative Low-Resolution Representations of Its Isocommittor Hypersurfaces.
Alessia Guadagnin Pattaro, Roberto Menichetti, Raffaello Potestio
Journal of chemical theory and computation1 Sep 2026Peer-reviewedCC BYFull textMolecular biology & gene regulationEurope PMC
Sequence design for three-dimensional genome folding using Akita Semifreddo
Paulina N. Smaruj, David R. Kelley, Geoffrey Fudenberg
Genome Biology26 Oct 2026Peer-reviewedCC BY-NC-ND 4.0Molecular biology & gene regulationCrossref
An inverse-protein-folding pre-trained model for predicting the mutational effects in protein-protein interactions
Sihong Mao, Siwei Meng, Ye Xiang
iScience1 Oct 2026Peer-reviewedCC BY 4.0Full textMolecular biology & gene regulationCrossrefEurope PMC
The interaction between NC(p7)1-55 and p6 may regulate interactions with nucleic acids during assembly through modulation of Gag folding.
V. LARUE, S. Nonin-Lecomte
bioRxiv28 Aug 2026PreprintAll rights reservedFull textBiophysics Structural biologybioRxivHAL
Cardiolipin regulates free energy of folding and vesicle leakage of CM15 antimicrobial peptides in bacterial inner membranes
N. M. Bahuguna, A. Upasani, A. Kulshrestha +3
bioRxiv5 Sep 2026PreprintAll rights reservedFull textBiophysicsbioRxiv

Cite

BibTeX RIS

BibTeX

@misc{jang2026cofolding,
  title = {{Co-folding with a Soup of Representations}},
  author = {Hyosoon Jang and Taewon Kim and Sungsoo Ahn},
  howpublished = {arXiv},
  year = {2026},
  month = sep,
  doi = {10.48550/arxiv.2609.15552},
  eprint = {2609.15552},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.15552},
  note = {CC BY 4.0},
}

RIS

TY  - UNPB
TI  - Co-folding with a Soup of Representations
AU  - Hyosoon Jang
AU  - Taewon Kim
AU  - Sungsoo Ahn
PY  - 2026
DA  - 2026/09/28
PB  - arXiv
DO  - 10.48550/arxiv.2609.15552
UR  - https://arxiv.org/abs/2609.15552
AB  - Co-folding models such as AlphaFold3, Protenix, ESMFold2, and OpenDDE have advanced rapidly, yet no single model consistently performs best across all biomolecular complexes. In this paper, we show that their pair representations encode complementary information that can be transferred across models to improve structure prediction. We introduce SoupFold, which combines pair representations from multiple co-folding models in a common representation space and generates structures from the combined representation. Importantly, SoupFold does not retrain the co-folding models and learns only simple mappings to transfer representations across models. We evaluate SoupFold on antibody-antigen, protein-protein, protein-ligand, molecular glue, GPCR, and oligomeric complex prediction using AlphaFold3, Protenix, ESMFold2, and OpenDDE. By combining representations across models, SoupFold improves over individual co-folding models across the considered benchmarks.
ER  -

Abstract

Related

Cite