Skip to content
DraconiticResearch
Search Sources API
203,203 studies31 sourcesUpdated 9 min ago
Peer-reviewedCC BYNature Methods

Aligning protein-generative models to experimental fitness with ProteinDPO

  • Talal Widatalla1,2
  • Ashir A. Borah1,3
  • S. B. King1,2
  • Claudia L. Driscoll1,2
  • Rafael Rafailov2
  • Brian Hie4,5
5 affiliations
  1. 1Arc Institute, Palo Alto, CA, USA
  2. 2Stanford University, Stanford, CA, USA
  3. 3University of California, San Francisco, San Francisco, CA, USA
  4. 4Arc Institute, Palo Alto, CA, USA. [email protected]
  5. 5Stanford University, Stanford, CA, USA. [email protected]
Open at sourcePDFDOICite

Abstract

Biological generative models can predict biological functions without task-specific training data but often under-perform specialized models. This is due to a fundamental 'alignment gap', where the rules learned during unsupervised training are not related to the function of interest. Here we demonstrate how to provide task-specific information without losing the general knowledge learned during pretraining by using direct preference optimization to align a structure-conditioned protein language model to preferentially generate stable protein sequences. Our aligned model, ProteinDPO, achieves stability prediction competitive to task-specific models and consistently outperforms unsupervised and fine-tuned versions of the model. Notably, ProteinDPO generalizes beyond its training data to enable stabilization and improved binding affinity prediction of large multichain protein complexes. When applied to stabilization of the hemagglutinin trimer, a primary component of influenza vaccines, ~80% of designs achieve increased or similar stability compared with the native hemagglutinin and up to 32 °C improvements from recently emerged mammalian strains. Our results demonstrate how to augment generative models with biophysical information and, more broadly, provide a general framework for the alignment of biological foundation models.

Related

  1. Understanding language model scaling for protein fitness prediction
    Chao Hou, Ди Лю, Aziz Zafar +1
    Nature Computational SciencePeer-reviewedLicence not statedFull text22 citedMolecular biology & gene regulationOpenAlex
  2. Representation learning for low-data protein fitness prediction using sequence–structure embeddings and graph neural networks
    Chanoknun Phunnasorn, Narupat Hongdilokkul, Wasakorn Laesanklang +2
    Scientific ReportsPeer-reviewedCC BY-NC-ND 4.0Full textMolecular biology & gene regulationCrossref
  3. Prospective discovery of LasB inhibitors by integrating generative design, free energy ranking, and experimental validation
    A. Martinez Leon, A. Vgenopoulou, G. Jezequel +3
    bioRxivPreprintCC BYFull textBiophysicsbioRxiv
  4. A small molecule increases mycobacterial translation fidelity without a fitness cost by targeting ribosomal protein S5
    Jie Wu, Swarnava Chaudhuri, Sarah Caldwell Feid +5
    mBioPeer-reviewedCC BY 4.0Full textMolecular biology & gene regulationCrossref
  5. The physical foundation beneath protein generative modeling
    Tianyu Lu, Po-Ssu Huang
    arXivPreprintCC BY 4.0Full textBiophysicsarXiv

Cite

BibTeXRIS
BibTeX
@article{widatalla2026aligning,
  title = {{Aligning protein-generative models to experimental fitness with ProteinDPO}},
  author = {Talal Widatalla and Ashir A. Borah and S. B. King and Claudia L. Driscoll and Rafael Rafailov and Brian Hie},
  journal = {Nature Methods},
  year = {2026},
  month = aug,
  doi = {10.1038/s41592-026-03137-3},
  url = {https://doi.org/10.1038/s41592-026-03137-3},
  note = {CC BY},
}
RIS
TY  - JOUR
TI  - Aligning protein-generative models to experimental fitness with ProteinDPO
AU  - Talal Widatalla
AU  - Ashir A. Borah
AU  - S. B. King
AU  - Claudia L. Driscoll
AU  - Rafael Rafailov
AU  - Brian Hie
PY  - 2026
DA  - 2026/08/14
JO  - Nature Methods
DO  - 10.1038/s41592-026-03137-3
UR  - https://doi.org/10.1038/s41592-026-03137-3
AB  - Biological generative models can predict biological functions without task-specific training data but often under-perform specialized models. This is due to a fundamental 'alignment gap', where the rules learned during unsupervised training are not related to the function of interest. Here we demonstrate how to provide task-specific information without losing the general knowledge learned during pretraining by using direct preference optimization to align a structure-conditioned protein language model to preferentially generate stable protein sequences. Our aligned model, ProteinDPO, achieves stability prediction competitive to task-specific models and consistently outperforms unsupervised and fine-tuned versions of the model. Notably, ProteinDPO generalizes beyond its training data to enable stabilization and improved binding affinity prediction of large multichain protein complexes. When applied to stabilization of the hemagglutinin trimer, a primary component of influenza vaccines, ~80% of designs achieve increased or similar stability compared with the native hemagglutinin and up to 32 °C improvements from recently emerged mammalian strains. Our results demonstrate how to augment generative models with biophysical information and, more broadly, provide a general framework for the alignment of biological foundation models.
SN  - 1548-7091
SN  - 1548-7105
ER  -