A structure-aware generative AI framework for revealing functional relationships in protein families
Proceedings of the National Academy of Sciences · 2026
DOI: 10.1073/pnas.2536965123 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the National Academy of Sciences, 2026. Verfasst von Divyanshu Shukla und 3 weiteren Autor:innen. Beteiligte Einrichtung: Bioinformatics and Computational Biology Program, Iowa State University, Ames, IA 50011. Hauptthemen laut MEDLINE: Proteins. Die Arbeit nennt 56 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: HHS | NIH | NIGMS | Native American Research Centers for Health (NARCH), NIGMS NIH HHS, NIAID NIH HHS.
Abstract (Originalsprache)
Proteins can be studied through their sequence statistics or structural properties. These represent complementary views that are useful but lack a quantitative framework to tell, family by family, which is most informative and how to combine them. We introduce a framework that builds both views in parallel: amino acid (AA) alignments are translated into parallel alignments over a 3D interaction (3Di) structure-informed alphabet. Variational autoencoders compress each into a two-dimensional map, and direct coupling analysis places a shared coevolutionary energy on both maps, turning them into latent generative landscapes. On these landscapes, we define information-theoretic distance metrics that quantify how sequence changes drive structural and functional variation in protein families. We demonstrate the framework on five families: in malate dehydrogenases, the 3Di landscape identifies the structurally conserved scaffold that this family uses to encode thermal adaptation via sequence variability revealed in the AA landscape. In globins and transient receptor potential melastatin (TRPM), the 3Di landscape recovers known functional subfamilies. In the Flaviviridae E1 and E2 glycoproteins, structure reveals evolutionary relationships invisible at the sequence level. Because many sequences encode the same fold, our framework lets us disentangle family-sequence variability from structural and functional variation. These generative landscapes allow sampling near functional regions, and we show they can help us gain mechanistic insight into the evolutionary forces shaping sequence–structure–function variation and guide the design of new proteins.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Stichwörter der Autor:innen
Substanzen
Autor:innen und Institutionen
- Bioinformatics and Computational Biology Program, Iowa State University, Ames, IA 50011.
- Department of Biological Sciences, The University of Texas at Dallas, Dallas, TX 75080.
- Department of Bioengineering, The University of Texas at Dallas, Dallas, TX 75080.
- Department of Physics, The University of Texas at Dallas, Dallas, TX 75080.
- Center for Systems Biology, The University of Texas at Dallas, Dallas, TX 75080.
- Department of Chemistry, Iowa State University, Ames, IA 50011.
ORCID-Profile: Divyanshu Shukla, Jonathan Martin, Faruck Morcos, Davit A Potoyan
Förderung
- HHS | NIH | NIGMS | Native American Research Centers for Health (NARCH)
- NIGMS NIH HHS
- NIAID NIH HHS
Zitieren
APA
Shukla, D., Martin, J., Morcos, F., & Potoyan, D. A. (2026). A structure-aware generative AI framework for revealing functional relationships in protein families. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2536965123
Vancouver
Shukla D, Martin J, Morcos F, Potoyan DA. A structure-aware generative AI framework for revealing functional relationships in protein families. Proceedings of the National Academy of Sciences. 2026. doi:10.1073/pnas.2536965123
BibTeX
@article{Shukla2026,
author = {Shukla, Divyanshu and Martin, Jonathan and Morcos, Faruck and Potoyan, Davit A},
title = {A structure-aware generative AI framework for revealing functional relationships in protein families},
journal = {Proceedings of the National Academy of Sciences},
year = {2026},
doi = {10.1073/pnas.2536965123}
}
EUVIMED