Convergence of cryo-electron microscopy and artificial intelligence in integrative structural biology: A critical review of advances, synergies, and implications for molecular biophysics and drug discovery

Abstract

Structural biology has entered a period of rapid methodological change defined by the convergence of cryo-electron microscopy (cryo-EM) and artificial intelligence (AI)-driven structure prediction. Prior reviews have generally addressed cryo-EM instrumentation or AI structure prediction individually, or their combination primarily from a structural biology or drug discovery perspective; a microscopy-centred synthesis tracing this convergence from detector physics and image-processing workflows through validation standards to in situ cryo-electron tomography (cryo-ET), while situating AI tools specifically as inputs to and complements of the cryo-EM workflow, has been comparatively underexplored. This review provides such a synthesis, organised around: (i) the instrumentation, image-processing, and validation advances underlying cryo-EM's resolution gains, including detector physics, contrast transfer function (CTF) estimation, Bayesian and deep-learning-based particle picking and reconstruction, and map-to-model validation and deposition; (ii) cryo-electron tomography and subtomogram averaging (STA) for in situ structural biology, including deep generative approaches to heterogeneity analysis; and (iii) the AlphaFold2, AlphaFold3, and RoseTTAFold All-Atom systems, considered specifically in relation to how they interface with and depend upon cryo-EM data for model building, refinement, and validation. Building on this microscopy-centred account, we propose a three-tier classification framework for selecting between AI-primed, experiment-led, and in situ integrative pipelines, and we summarise performance, transparency, and accessibility differences between leading AI tools. We illustrate these methods with examples from membrane proteins, ion channels, and viral glycoproteins, and briefly note implications for structure-based drug design and for structural biology capacity in resource-limited settings. Throughout, we distinguish between demonstrated capability, exceptional proof-of-principle results, and routine practice, and identify the experimental and computational limitations that constrain each approach.