A quantitative approach to characterise model misspecification in cryo-EM

Simulation-based inference can significantly speed up single-particle cryo-EM analysis. However, simulated data often doesn’t accurately compare with experimental data, and the mismatch isn’t well-defined. We recast the problem of characterising this model misspecification as a geometric one: we embed experimental and synthetic images in the latent space of a foundation model trained on experimental images and measure the misalignment between their distributions. We use these measurements to help admit and fine-tune the image-formation terms of a forward model; we build a differentiable, physics-informed simulator, which can optimise its hyperparameters towards a latent-space objective. Across seven experimental datasets, our simulated distribution covers more of the experimental distributions and aligns better with them than an established simulation baseline. A foundation model’s latent space, along with an experimental image stack, can then serve as a measuring instrument for the simulation-to-experiment gap.