Skip to content
DraconiticResearch
Search Sources API
197,688 studies31 sourcesUpdated 31 min ago
PreprintCC BYbioRxiv

How do Co-Folding Models Organize Structural Information?

  1. 1Department of Chemistry, KAIST

Abstract

Co-folding models accurately predict biomolecular complexes, but how their internal representations support joint structure prediction remains unclear. We analyze the Boltz-1 trunk by decomposing its pair representation into intra-chain and inter-chain blocks and applying ablation, layer-wise representation geometry, and activation patching. Our results show that structural information is distributed across three streams: the single representation carries token-internal atomic detail, the intra-chain pair representation encodes chain-level geometry, and the inter-chain pair representation controls relative molecular arrangement. Across the trunk depth, representations follow a Mix-Compress-Refine trajectory, while linear probes reveal that intra-chain distance content is largely supplied by MSA preconditioning and inter-chain distance content is constructed progressively across Pairformer layers at each recycle. Apo-holo patching further shows that ligand-induced backbone changes are not fully resolved before diffusion; instead, the diffusion module reconciles partially inconsistent intra- and inter-chain constraints. These findings provide a mechanistic view of how co-folding trunks organize structural information and suggest more partition-aware designs for future biomolecular structure prediction models.

Related

  1. Filling the Missing Puzzle: Integrating Interaction and Evolutionary Information to Enhance Protein Fitness Prediction in Inverse Folding
    Z. Wang, Z. Zhang, Y. Yang +6
    bioRxivPreprintCC BY-NCFull textBioinformatics & computational biologybioRxiv
  2. Frequency-specific oscillations track parallel processing of linguistic structural and lexical-semantic information from perception to decision
    Yanlin Fu, Zhao Xu, Qinhui Yao +2
    Communications BiologyPeer-reviewedCC BY-NC-ND 4.0Full textNeuroscienceCrossref
  3. Mitotic centromeres organize in repeat-subunits with specialized histone landscape and multi-domain architecture as revealed by expansion microscopy
    G. Thangavel, S. Erhardt
    bioRxivPreprintAll rights reservedFull textCell biologybioRxiv
  4. Artificial intelligence-based decision support for axial spondyloarthritis integrating inflammatory and structural MRI information
    Yunseo Kim, Seulkee Lee, Seonyoung Kang +5
    Frontiers in ImmunologyPeer-reviewedCC BY 4.0Full textImmunologyCrossrefEurope PMCPMCEurope PMC
  5. Cellular source and circuit context organize functional specificity in the Drosophila NPF system
    M. N. Endres, T. S. Dadyala, K. W. Christie +2
    bioRxivPreprintCC BYFull text1 citedNeurosciencebioRxiv
BibTeX
@misc{park2026cofolding,
  title = {{How do Co-Folding Models Organize Structural Information?}},
  author = {M. Park and S. Kim and S. Moon and H. Kim and G. Jeon and W. Y. Kim},
  howpublished = {bioRxiv},
  year = {2026},
  month = sep,
  doi = {10.64898/2026.09.06.749739},
  url = {https://www.biorxiv.org/content/10.64898/2026.09.06.749739v1},
  note = {CC BY},
}
RIS
TY  - UNPB
TI  - How do Co-Folding Models Organize Structural Information?
AU  - M. Park
AU  - S. Kim
AU  - S. Moon
AU  - H. Kim
AU  - G. Jeon
AU  - W. Y. Kim
PY  - 2026
DA  - 2026/09/06
PB  - bioRxiv
DO  - 10.64898/2026.09.06.749739
UR  - https://www.biorxiv.org/content/10.64898/2026.09.06.749739v1
AB  - Co-folding models accurately predict biomolecular complexes, but how their internal representations support joint structure prediction remains unclear. We analyze the Boltz-1 trunk by decomposing its pair representation into intra-chain and inter-chain blocks and applying ablation, layer-wise representation geometry, and activation patching. Our results show that structural information is distributed across three streams: the single representation carries token-internal atomic detail, the intra-chain pair representation encodes chain-level geometry, and the inter-chain pair representation controls relative molecular arrangement. Across the trunk depth, representations follow a Mix-Compress-Refine trajectory, while linear probes reveal that intra-chain distance content is largely supplied by MSA preconditioning and inter-chain distance content is constructed progressively across Pairformer layers at each recycle. Apo-holo patching further shows that ligand-induced backbone changes are not fully resolved before diffusion; instead, the diffusion module reconciles partially inconsistent intra- and inter-chain constraints. These findings provide a mechanistic view of how co-folding trunks organize structural information and suggest more partition-aware designs for future biomolecular structure prediction models.
ER  -