Cryo‑EM model building gets faster

Cryo-electron microscopy works by flash-freezing proteins, taking huge numbers of two-dimensional pictures with an electron beam, and combining them into a three-dimensional density map that looks more like a foggy statue than a finished molecule. Structural biologists then have to turn that map into an atomic model, placing each amino acid where it belongs. (nature.com) That last step has been one of the slowest parts of the whole workflow. A 2025 review in Current Opinion in Structural Biology says model building still often depends on experts manually tracing chains through density maps in graphics software, especially when the map is incomplete or noisy. (sciencedirect.com) Deep learning is changing that because density maps are image-like data, and image-recognition systems are good at spotting patterns humans used to trace by hand. The review says recent tools now detect helices, sheets, backbone paths, and even full atomic arrangements directly from the map. (sciencedirect.com) The field’s big shift is from “find one feature at a time” to “build the whole structure in one pass.” The review groups the new methods into de novo systems, which predict a model straight from the density, and hybrid systems, which combine the map with outside structure predictions or templates. (frontiersin.org) One of the clearest examples is ModelAngelo, published in Nature in 2024. Its authors wrote that interpreting cryo-electron microscopy maps usually requires “labour-intensive manual intervention,” and showed that their machine-learning system could automatically build atomic models and identify unknown proteins from the map itself. (nature.com) That matters because cryo-electron microscopy is producing structures faster every year. The ModelAngelo paper notes that the number of new entries in the Electron Microscopy Data Bank has been growing exponentially, and projected that roughly 100,000 cryo-electron microscopy structures could be determined within 5 years if the trend continued. (nature.com) Newer systems are pushing the automation further. A 2025 Nature Structural & Molecular Biology paper on CryoAtom reported more complete models than ModelAngelo, lower resolution requirements, and faster modeling by adapting AlphaFold-style architecture to local three-dimensional map information. (nature.com) The review says these tools are most useful when the map is good enough to show the protein’s backbone but not clean enough for a person to read every atom quickly. That is the structural biology version of switching from tracing roads by hand on a blurry satellite image to having software draw the street map for you. (sciencedirect.com) The bottleneck has not disappeared. The review says nucleic acids, mixed protein-nucleic acid assemblies, flexible regions, and lower-resolution maps are still harder for deep-learning systems than well-behaved protein maps, so human checking is still part of the job. (sciencedirect.com; academic.oup.com) But the direction is clear: more of the “build the model” step is moving from artisanal work to software. A 2025 benchmark and survey in Briefings in Bioinformatics found that deep-learning methods now form a distinct class of model-building tools, especially when paired with sequence-based predictors such as AlphaFold, and are being measured directly against older physics-based pipelines. (academic.oup.com) When that handoff works, the payoff is not abstract. Faster model building means a lab can move sooner from a frozen sample to the actual shape of a drug target, a viral machine, or a protein complex, and spend less time on the digital equivalent of connecting dots one atom at a time. (sciencedirect.com; nature.com)