From generative protein models to experimental proteins: Validation and benchmarking for deep generative protein design
Artificial Intelligence in the Life Sciences · 2026-11 · 100191
DOI: 10.1016/j.ailsci.2026.100191 ↗Kurz gefasst
Fachartikel in der Zeitschrift Artificial Intelligence in the Life Sciences, 2026-11. Verfasst von Wenran Li und 9 weiteren Autor:innen. Die Arbeit nennt 59 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Recent advances in generative artificial intelligence have shifted computational protein engineering from structure prediction toward de novo biomolecular design. Diffusion- and flow-based approaches are especially prominent because they can generate diverse protein structures under geometric and functional constraints. However, heterogeneity in molecular representations, conditioning strategies, evaluation pipelines, and definitions of success complicates comparisons across studies. This review examines deep generative protein design from candidate generation to experimental validation. We summarize major representations, generative frameworks, and conditioning strategies for de novo backbone generation, motif scaffolding, binder design, enzyme design, and multimolecular generation. We organize validation as a hierarchy spanning computational filtering, synthesis and expression, biophysical characterization, structural confirmation, and functional testing. A matched-target PD-L1 binder-design case study comparing RFdiffusion and BindCraft illustrates substantial pre-experimental attrition. The case study illustrates workflow-dependent computational attrition, while retrospective analysis demonstrates sensitivity to final filtering thresholds. We therefore distinguish stage-specific computational success from experimental and global yield and emphasize that every success rate should specify its numerator and denominator. We also examine benchmarking bias and uncertainty arising from circular evaluation, evaluator dependence, dataset leakage, generation budget, physics-based modeling, and experimental variability, while treating failed candidates as informative outcomes. Finally, we propose minimum reporting standards that make attrition reconstructable across the complete design pipeline. This framework supports more reproducible and interpretable benchmarking and strengthens the connection between generative models and experimentally validated proteins.
Zitieren
APA
Li, W., Cadet, X., Medina-Ortiz, D., Charton, P., Davari, M. D., Li, Y., Damour, C., Miranville, A., Brevern, A. D., & Cadet, F. (2026). From generative protein models to experimental proteins: Validation and benchmarking for deep generative protein design. Artificial Intelligence in the Life Sciences, 100191. https://doi.org/10.1016/j.ailsci.2026.100191
Vancouver
Li W, Cadet X, Medina-Ortiz D, Charton P, Davari MD, Li Y, et al. From generative protein models to experimental proteins: Validation and benchmarking for deep generative protein design. Artificial Intelligence in the Life Sciences. 2026:100191. doi:10.1016/j.ailsci.2026.100191
BibTeX
@article{Li2026,
author = {Li, Wenran and Cadet, Xavier and Medina-Ortiz, David and Charton, Philippe and Davari, Mehdi D. and Li, Yu and Damour, Cedric and Miranville, Alain and Brevern, Alexandre de and Cadet, Frederic},
title = {From generative protein models to experimental proteins: Validation and benchmarking for deep generative protein design},
journal = {Artificial Intelligence in the Life Sciences},
year = {2026},
pages = {100191},
doi = {10.1016/j.ailsci.2026.100191}
}
EUVIMED