AI tools for structural biology can find 'structure' in pure noise

AI tools used in structural biology to identify particles can create false structures from random noise. This means that researchers using these tools might be seeing patterns that are not actually there, especially in low-quality data.

Scientists use cryo-electron microscopy to map the 3D structures of molecules, which is critical for drug discovery and understanding diseases. If the initial AI steps in this process are biased, the resulting 3D models could be wrong. This paper shows that the bias is not just a possibility, but a mathematical certainty under certain conditions, and it can be reproduced with standard software.