AI Takes Over Cryo-EM: Deep Neural Networks Now Automate Particle Picking, Map Reconstruction and Atomic Modelling

Artificial intelligence is reshaping cryo-electron microscopy, the technique that lets structural biologists determine macromolecular structures at near-atomic resolution. A comprehensive review published in BMC Artificial Intelligence examines how deep neural networks are redefining every stage of the cryo-EM workflow, from automated particle picking and map reconstruction to structure modelling and high-resolution refinement. The authors, led by Saleem Iqbal and Bairong Shen, argue that the integration of deep learning and neural network-based tools has transformed cryo-EM into a data-driven and highly precise discovery platform, positioning AI as an integral component of the so-called resolution revolution in structural biology.

The timing is significant. In 2024, John Hopfield and Geoffrey Hinton received the Nobel Prize in Physics for computational methodologies that underpin the current surge in AI research, while David Baker, Demis Hassabis and John Jumper shared the Chemistry prize for computational tools that predict and design protein structures. Cryo-EM itself earned a Nobel in 2017, when Jacques Dubochet, Richard Henderson and Joachim Frank were recognised for developing the technique for high-resolution structure determination of biomolecules in solution. The convergence of these two prize-winning fields is now accelerating molecular insights with implications for structural biology and drug discovery.

Cryo-EM captures 2D images of biological macromolecules and assemblies, such as protein particles and viruses, at cryogenic temperatures, using direct electron detectors, motion correction algorithms and Bayesian reconstruction methods. Its ability to capture multiple conformational states provides mechanistic insights critical for understanding molecular function. Yet traditional data processing pipelines struggle with large, heterogeneous datasets. Manual particle selection, extensive computational demands and uneven sampling of conformational states often limit resolution, especially for rare or transient conformations. Deep neural networks now facilitate automated particle picking, denoising, classification and three-dimensional reconstruction, substantially reducing human bias and computational bottlenecks.

The workflow begins with sample preparation, which directly impacts the quality of the resulting cryo-EM map. Factors such as pH, salt concentration, storage conditions, sample concentration and purification steps significantly influence single-particle quality. Classical methods including gel electrophoresis, size exclusion chromatography and dynamic light scattering are used to assess purity, heterogeneity, aggregation and complex stability, while negative-staining transmission electron microscopy offers a rapid means to visually estimate particle distribution on grids. A major concern is the complex interaction between samples suspended in liquid and the surrounding atmosphere, particularly at the air-water interface, where particle denaturation and orientation bias persist and no universally applicable solution has yet been developed. AI tools such as VitrifAI aim to optimise vitrification and potentially improve particle quality, although AI is not directly involved in the physical process of sample preparation or grid optimisation.

Particle picking remains a critical bottleneck. Individual protein particles must be identified in micrographs before structures can be reconstructed, but the standard template-based method is labour intensive and time consuming, often selecting invalid particles such as false positives amid ice contamination, carbon areas and overlapping particles. Supervised methods such as DeepPicker and crYOLO rely on labelled datasets like CryoPPP to train convolutional neural networks for precise identification. Unsupervised techniques, exemplified by AutoCryoPicker, bypass labelled data by employing clustering algorithms to detect particles based on intrinsic structural features. Semi-supervised frameworks bridge the gap by combining supervised architectures with foundational models, while adversarial approaches such as CryoGAN use generative networks to iteratively refine particle localisation and map quality.

One of the standout tools is CryoSegNet, which integrates attention-gated U-Nets with Meta’s Segment Anything Model. The specialised U-Net, trained on the CryoPPP dataset, produces a segmentation map that serves as input for SAM’s automatic mask generator, enabling precise localisation of protein particles even in low signal-to-noise micrographs. Applying SAM directly to cryo-EM images segments only a limited number of particles, because cryo-EM images differ significantly from SAM’s training data, and retraining SAM on cryo-EM images produced only marginal improvements. The hybrid strategy proved decisive: CryoSegNet outperformed the two leading AI-based pickers, crYOLO and Topaz, in both particle picking accuracy and the resolution of 3D protein density maps reconstructed from the identified particles.

The numbers are striking. Reconstructed density maps built from particles selected by CryoSegNet achieved an average resolution of 3.33 angstroms, 7 percent better than Topaz’s 3.58 angstroms and 14 percent better than crYOLO’s 3.87 angstroms. The tool also excelled in key image segmentation metrics such as the F1 score, precision and Dice score, all of which strongly correlate with the resolution of the reconstructed density maps. Regardless of the kind, shape or size of the protein, CryoSegNet exhibited strong recall and precision in segmenting protein particles from cryo-EM micrographs, and it surpassed SAM and the two leading machine learning pickers on many independent datasets of different protein types.

Map reconstruction and modelling have been transformed in parallel. CryoDRGN2 enables ab initio neural reconstruction of 3D protein structures directly from cryo-EM images, while CryoDRGN-AI achieves ab initio reconstruction of heterogeneous complexes, combining neural representations with exhaustive search strategies to resolve previously inaccessible conformational states from raw data. For model building, DeepTracer and ModelAngelo employed experimental density maps from the Electron Microscopy Data Bank, moving beyond earlier approaches that depended on simulated maps derived from Protein Data Bank structures, which frequently failed to capture the intricacies of actual data. DeepMainmast integrates geometric deep learning for domain-level assembly, and EmodelX resolves conformational landscapes from heterogeneous data. In postprocessing, DeepEMhancer and EMReady refine low-resolution maps, although leveraging large datasets can introduce model bias for structures underrepresented in the data.

A newer technique, CryoTRANS, addresses the shortcomings of deep learning post-processing methods by modelling the spectrum of nonrigid transformations between prevalent and infrequent conformational states. It utilises a deep neural network formulated as an ordinary differential equation, adopting a self-supervised learning approach that eliminates the need for supervised samples or pretraining, thus minimising model bias. Applied to four real cryo-EM datasets, it exhibited superior performance in predicting high-resolution maps, outperforming DeepEMhancer and EMReady. In cryo-electron tomography, AI methods are now essential for addressing tomographic noise, missing-wedge artifacts and complex cellular backgrounds, with tools such as DeepFinder, IsoNet and the open-source AlTom platform enabling increasingly automated analysis of cellular landscapes at molecular resolution.

Challenges persist. Data scarcity remains a formidable barrier, since most existing machine learning methods for particle picking are trained and tested on only a few manually labelled datasets of a few proteins, such as Apoferritin and Keyhole Limpet Hemocyanin, and cannot generalise well to the diverse, irregular protein shapes generated daily by cryo-EM facilities worldwide. The detection of rare conformational states is hindered by the scarcity of particle images, and the computational expense of training models such as CryoSegNet presents a barrier for smaller laboratories. The Cryo2Structure dataset predominantly includes density maps with resolutions below 4 angstroms, underscoring a scarcity of publicly accessible datasets for medium to low-resolution maps. The review’s authors call for lightweight AI architectures, expanded datasets covering underrepresented resolution ranges such as 5 to 8 angstroms, and end-to-end deep learning frameworks akin to AlphaFold. Cryo-EM remains unsuitable for intrinsically disordered proteins lacking stable structural states, necessitating integrative approaches with NMR or AlphaFold. Still, the synergy between AI and cryo-EM continues to push the boundaries of computational biology, offering deeper mechanistic insights and tangible benefits for biomedical research and human health.

Subject of Research: Deep learning methods across the cryo-EM structural determination workflow

Article Title: Artificial intelligence in cryo-EM: emerging deep neural network methods from sample preparation, particle picking, map reconstruction, modelling to enhanced resolution

Article References: Iqbal, S., Eng, E. T., Kamal, M. A., & Shen, B. (2026). Artificial intelligence in cryo-EM: emerging deep neural network methods from sample preparation, particle picking, map reconstruction, modelling to enhanced resolution. BMC Artificial Intelligence, 2(1), Article 2. https://doi.org/10.1186/s44398-025-00017-2

Image Credits: AI Generated

DOI: 10.1186/s44398-025-00017-2

Keywords: cryo-EM, deep learning, particle picking, map reconstruction, structural biology, CryoSegNet, CryoDRGN, ModelAngelo, AlphaFold, drug discovery, neural networks, resolution revolution