Structure Prediction of Large RNAs with AlphaFold3 Highlights its Capabilities and Limitations.

ABSTRACT

DeepMind's AlphaFold3 webserver offers exciting new opportunities to make structural predictions of heterogeneous macromolecular systems. Here we attempt to apply AlphaFold3 to large RNA molecules whose 3D atomic structures are unknown but whose physical dimensions have been studied experimentally. One difficulty that we encounter is that models returned by AlphaFold3 often contain severe steric clashes and, less frequently, clear breaks in the phosphodiester backbone, with the probability of both events increasing with the length of the RNA. Restricting attention to those RNAs for which non-clashing models can be obtained, we find that hydrodynamic radii computed from the AlphaFold3 models are much larger than those reported experimentally under low salt conditions but are in better agreement with those reported in the presence of polyvalent cations. For two RNAs whose shapes have been imaged experimentally, the computed anisotropies of the AlphaFold3-predicted structures are too low, indicating that they are excessively spherical; extending this analysis to larger RNAs shows that they become progressively more spherical with increasing length. Overall, the results suggest that AlphaFold3 is capable of producing plausible models for RNAs up to ∼2000 nucleotides in length, but that thousands of predictions may be required to obtain models free of geometric problems.