Artificial intelligence for the identification and structure determination of macromolecules in cryo-electron microscopy and tomography

Abstract

Cryo-electron microscopy (cryo-EM) and cryo-electron tomography (cryo-ET) are powerful technologies for determining the three-dimensional structures of biological macromolecules. However, particle identification, a crucial step in these workflows, remains challenging due to limited annotated training data, low signal-to-noise ratios, particle heterogeneity, background artifacts in cryo-EM, and missing-wedge effects and crowded environments in cryo-ET. This dissertation addresses these challenges through four major contributions. First, a comprehensive review and comparative evaluation of AI-based cryo-EM particle-picking methods was conducted, and existing methods were benchmarked using biologically meaningful metrics, including reconstructed 3D density map resolution and viewing-direction coverage, to assess their strengths and identify directions for future development. Second, we curated CryoPPP, a large open-access dataset of 9,893 manually labeled micrographs across 34 diverse proteins, comprising 2.7 million annotated particles validated against gold-standard labels. Third, we developed CryoTransformer, an end-to-end transformer-based particle picker that achieved state-of-the-art performance, outperforming the existing methods. Finally, for cryo-ET, we developed TomoSwin3D, a 3D Swin Transformer U-Net that achieved state-of-the-art macromolecule detection across several independent benchmarks, with particularly strong gains on small and low-contrast targets. Together, these contributions establish a rigorous foundation for AI-driven macromolecule identification in structural biology, with direct implications for high-resolution structure determination, visual proteomics, drug discovery, and vaccine design.