Synthetic sequence alignments as programmable probes of learned conformational landscapes in deep learning protein structure predictors
Bioinformatics · 2026
DOI: 10.1093/bioinformatics/btag605 ↗Kurz gefasst
Fachartikel in der Zeitschrift Bioinformatics, 2026. Verfasst von Jannik Adrian Gut und 2 weiteren Autor:innen. Beteiligte Einrichtung: Institute of Biochemistry and Molecular Medicine, University of Bern, Bern 3012, Switzerland. Hauptthemen laut MEDLINE: Deep Learning, Proteins, Sequence Alignment, Computational Biology. Die Arbeit nennt 61 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: FreeNovation 2023, Swiss National Science Foundation.
Abstract (Originalsprache)
Motivation Proteins rely on conformational flexibility for biological function, yet predicting alternative states remains a major challenge in structural biology. Although deep learning models like AlphaFold2, AlphaFold3, and RoseTTAFold2 excel at static structure prediction, what these networks actually learn about the underlying conformational landscapes remains largely opaque. Results Here we introduce synthetic multiple sequence alignments (MSAs), designed by inverse folding to encode predefined structural constraints, as a programmable intervention for interrogating the internal logic of structure prediction systems. Synthetic MSAs systematically bias AlphaFold2, AlphaFold3, and RoseTTAFold2 toward distinct conformational states of fold-switching proteins, including alternative conformations inaccessible through natural sequence information alone. Adversarial experiments pairing query sequences with MSAs encoding competing folds reveal sequence-dependent responses, exposing how alignment-derived and sequence-derived signals are weighted within each system. Probing predictions initialized from molecular dynamics trajectories reveals a systematic bias toward compact, training-distribution-favored conformations. Hybrid alignments combining synthetic and natural MSA segments enable targeted steering toward specific conformational states. These results suggest synthetic MSAs as a generalizable framework for dissecting the conformational landscapes encoded by deep learning structure predictors, with direct implications for understanding model behavior and accessing biologically relevant hidden states. Availability and implementation Newly generated data can be found at https://zenodo.org/records/20916910. The code underlying this article is available on GitHub at https://github.com/ibmm-unibe-ch/msa-tests.
Schlagwörter
MeSH-Schlagwörter (MEDLINE)
Substanzen
Autor:innen und Institutionen
- Institute of Biochemistry and Molecular Medicine, University of Bern, Bern 3012, Switzerland.
- Graduate School for Cellular and Biomedical Sciences (GCB), University of Bern, Bern 3012, Switzerland.
ORCID-Profile: Jannik Adrian Gut, Noah Kleinschmidt, Thomas Lemmin
Förderung
- FreeNovation 2023
- Swiss National Science Foundation
Zitieren
APA
Gut, J. A., Kleinschmidt, N., & Lemmin, T. (2026). Synthetic sequence alignments as programmable probes of learned conformational landscapes in deep learning protein structure predictors. Bioinformatics. https://doi.org/10.1093/bioinformatics/btag605
Vancouver
Gut JA, Kleinschmidt N, Lemmin T. Synthetic sequence alignments as programmable probes of learned conformational landscapes in deep learning protein structure predictors. Bioinformatics. 2026. doi:10.1093/bioinformatics/btag605
BibTeX
@article{Gut2026,
author = {Gut, Jannik Adrian and Kleinschmidt, Noah and Lemmin, Thomas},
title = {Synthetic sequence alignments as programmable probes of learned conformational landscapes in deep learning protein structure predictors},
journal = {Bioinformatics},
year = {2026},
doi = {10.1093/bioinformatics/btag605}
}
EUVIMED