Critical assessment of the quality of AlphaFold3 structural predictions for RNA and DNA: the imitation game

Published

March 30, 2025

Author(s)

Christina Bergonzo, Alexander Grishaev

Abstract

AlphaFold3 (AF3), available as a webserver, predicts the structures of all life's biomolecules. We evaluate its accuracy for oligonucleotides via agreement with experimental observables, as determination of RNA structure is known to present challenges for both experiment and computation. Although AlphaFold performs well for proteins, AF3 oligonucleotide predictions generally fall short of ground truth, especially for internal or capping loops, non-canonical base pairings, and cases involving conformational flexibility, all essential for RNA folding, interactions, and function. We observe r.m.s. errors in AF3 nucleotide orientations ranging between 7⁰ and 30⁰, with higher accuracies for simpler architectures of individual canonically-paired helical stems. With these mixed results, AF3 cannot be considered as a replacement for experimental RNA structure determination, highlighting the necessity of experimental validation of its predictions. Taking AlphaFold as a prime example of artificial intelligence tools, we observe imitation of the training dataset rather than reproduction of the underlying reality.

Citation

Journal of Chemical Information and Modeling

Pub Type

Journals

Download Paper

https://doi.org/10.1021/acs.jcim.5c00245

Local Download

Keywords

RNA structure, AlphaFold, artificial intelligence, AI, NMR, RDC

Citation

Bergonzo, C. and Grishaev, A. (2025), Critical assessment of the quality of AlphaFold3 structural predictions for RNA and DNA: the imitation game, Journal of Chemical Information and Modeling, [online], https://doi.org/10.1021/acs.jcim.5c00245, https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=958615 (Accessed April 2, 2025)