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

Area of Science:

• Structural biology
• Computational biology
• Biophysics

Background:

• Accurate prediction of macromolecular structures is crucial for understanding biological function.
• DeepMind's AlphaFold3 offers advanced capabilities for predicting the 3D atomic structures of complex biological molecules.
• Experimental studies of large RNA molecules provide physical dimensions that can be used to validate computational models.

Purpose of the Study:

• To evaluate the performance of AlphaFold3 in predicting the structures of large RNA molecules.
• To identify limitations and potential improvements for AlphaFold3 when applied to RNA.
• To compare AlphaFold3-derived structural properties with experimental data for RNA.

Main Methods:

• Application of the AlphaFold3 webserver to predict structures of large RNA molecules.
• Analysis of predicted models for geometric problems such as steric clashes and backbone breaks.
• Comparison of computed hydrodynamic radii and anisotropies from AlphaFold3 models with experimental data under varying salt conditions.
• Assessment of model quality as a function of RNA length.

Main Results:

• AlphaFold3 predictions for large RNAs frequently exhibit steric clashes and phosphodiester backbone breaks, with increasing probability for longer molecules.
• Hydrodynamic radii from non-clashing AlphaFold3 models are larger than experimental values under low salt but agree better with polyvalent cation conditions.
• Computed anisotropies suggest AlphaFold3 models are more spherical than experimentally observed for some RNAs, with increasing sphericity for longer molecules.
• Geometric problems in AlphaFold3 models increase with RNA length, necessitating multiple predictions.

Conclusions:

• AlphaFold3 can generate plausible models for RNAs up to approximately 2000 nucleotides.
• Obtaining geometrically valid RNA models from AlphaFold3 may require generating thousands of predictions.
• Further refinement or specialized approaches may be needed to improve the accuracy of AlphaFold3 predictions for large RNA structures, particularly regarding their shape and flexibility.