Realism and resolution: How AI-predicted protein structures differ from experimental structures

WhatRes shows that AI-predicted protein models can match apparent resolution yet exhibit systematically less variation in bond lengths and angles than high-resolution crystal structures.

• Why it matters: Predicted structures are often treated like experimental ones, but geometry-faithfulness is hard to judge from coordinates alone. WhatRes addresses this gap by separating “resolution-like” similarity from realism of geometry.
• What they did: The method evaluates structures on two axes: apparent resolution, inferred from local geometry similarity to experimental structures, and realism, measuring how closely that geometry matches experimental structures at the same apparent resolution.
• The result: Even high-confidence predictions reveal over-idealisation, indicating that good apparent resolution scores can mask reduced geometric variability versus high-resolution crystal structures.

AI summary of the abstract · may contain errors · not medical advice · read the paper.