structural-biology/ai in a supporting role/Nature 2022 · v2
Cryo-EM maps show water shifting in actin filaments as they age
Researchers determined six structures of actin filaments in different chemical states at resolutions of 2.15–2.24 Å. A neural network located the filaments in the microscope images; everything downstream was conventional processing and hand-built modelling.
spectrum · one line per step, placed by what the step does · bright lines used AI
Structural basis of actin filament assembly and aging
Nature, 2022
doi:10.1038/s41586-022-05241-8 · record aix-00029 v2 · checked 2026-10-07
- AI was for
- Detection
- Model family
- Convolutional neural network
- Checked by
- None stated
- Code
- not reported
AI processed or interpreted data, but the main finding does not rest on it.
What this research was about
Actin is one of the most abundant proteins in animal cells. Single actin molecules join end to end into long helical filaments, which give cells their shape and let them move, divide and pull on their surroundings. Each actin molecule carries a small fuel-like molecule, ATP, in a pocket. Once the molecule joins a filament, the ATP is cut, leaving ADP and a loose phosphate, which later escapes. That slow chemical change is how a filament 'ages', and the cell reads the age to decide which filaments to keep and which to dismantle. Seeing it means resolving not just the protein but the individual water molecules around the pocket.
To do that, the researchers used cryo-electron microscopy, in which samples are frozen and imaged with electrons, then thousands of images are averaged into a three-dimensional map. They prepared filaments held in three defined nucleotide states, with either magnesium or calcium, and solved six structures. They also tested how readily cofilin-1, a protein that cuts filaments, severed each state.
Where AI came in
One learned component appears in the work, early in the image-processing chain. After the raw films were corrected for drift and lens effects, a neural-network picker called SPHIRE_crYOLO was run to find the filaments in each micrograph and mark boxes along them, spaced 40 pixels per 27.8 Å apart. Those boxes became the particle set for everything that followed. The task it performed is detection: deciding where in a noisy image a filament lies, work that otherwise falls to a person tracing filaments by eye.
The tool was used as supplied, and the paper does not state its architecture, what it was trained on, or how accurately it picked. Nothing learned touched the rest of the study. The three-dimensional maps came from conventional helical refinement, and the atomic models, including every ion and water molecule, were placed and adjusted by hand. The cofilin severing experiments involved no model at all. The structural conclusions therefore do not rest on the network.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The authors determined six cryo-EM structures of rabbit skeletal alpha-actin filaments at resolutions of 2.15-2.24 Å, covering the ADP-BeF3-, ADP-Pi and ADP nucleotide states with either Mg2+ or Ca2+ bound. A neural-network filament picker was used to extract actin segments from the micrographs before conventional helical refinement and manual atomic model building. The structures resolve water molecules in the nucleotide-binding pocket, show that the pocket's water arrangement changes on polymerization, and show the proposed back door for phosphate exit closed in both the ADP-Pi and ADP states. Cofilin-1 severing assays found substantial severing of the ADP state but not the ADP-BeF3- or ADP-Pi states for Ca2+-F-actin, as for Mg2+-F-actin.
How AI was used
Machine learning entered this study at a single point in the cryo-EM image-processing pipeline. After on-the-fly preprocessing in TranSPHIRE, in which super-resolution videos were binned, gain and motion corrected with MotionCor2 and CTF-estimated with CTFFIND4.13, F-actin segments were located in the micrographs by the filament-picking procedure of SPHIRE_crYOLO, run with a box distance of 40 pixels per 27.8 Å and a minimum of six boxes per filament. The resulting boxes were extracted into 384 x 384 pixel particles and passed to non-learned steps: 2D classification with ISAC2, manual inspection and discarding of non-filament and ice classes, 3D helical refinement with meridien alpha in SPHIRE, Bayesian polishing and CTF refinement in Relion, 3D classification without alignment, and masked local refinement. Atomic models, including all ions and water molecules, were built and adjusted manually in Coot and refined with phenix real-space refine. The paper does not state the picker's architecture, training data or any evaluation of its picking accuracy, and no learned model contributes to the structural interpretation or to the cofilin severing assays.
The shape of the work
Structural · the record, drawn
no AI
Purify actin and reconstitute filaments in defined nucleotide states
Physical execution, by hand or by robot.
Actin polymerization was induced by the addition of 100 mM KCl and 2 mM CaCl2/MgCl2 (final concentrations).where the paper describes this · verbatim
no AI
Collect cryo-EM movie datasets
Obtaining raw data, whether by measurement, download or retrieval.
All datasets were collected on a 300 kV Titan Krios microscope (Thermo Fisher Scientific) equipped with a K3 detector (Gatan)where the paper describes this · verbatim
no AI
Preprocess movies: motion correction and CTF estimation
Cleaning, filtering, normalising or labelling data already obtained.
gain corrected and motion corrected using UCSF MotionCor2 (ref. ), contrast transfer function (CTF) estimations were performed with CTFFIND4.13where the paper describes this · verbatim
AI
Pick filament segments with a neural-network picker
Running a trained model over new data to predict, classify or score. The AI stood in for manual curation.
F-actin segments were picked using the filament picking procedure in SPHIRE_crYOLO using a box distance of 40 pixels per 27.8 Åwhere the paper describes this · verbatim
no AI
Classify and clean the particle set
Reducing a candidate set by filtering or ranking, in a single pass.
the particles were two-dimensionally classified in batches of 20,000 particles using ISAC2where the paper describes this · verbatim
no AI
Refine three-dimensional helical reconstructions
Iterative search over a space.
the particles were subjected to three-dimensional helical refinement using meridien alphawhere the paper describes this · verbatim
no AI
Build and refine atomic models including solvent
Extracting understanding from model behaviour.
All solvent molecules (ions and water molecules) were placed manually in Coot in the central actin subunitwhere the paper describes this · verbatim
no AI
Test cofilin severing across nucleotide states
Physical execution, by hand or by robot.
Severing assays were performed in 20 μl volumes by incubating 5 μM of F-actin with 5, 10 or 20 μM of cofilinwhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The only learned model in the workflow is the neural-network particle picker used to extract filament segments from micrographs; the structural conclusions come from conventional helical refinement and manual model building downstream of it.
F-actin segments were picked using the filament picking procedure in SPHIRE_crYOLO using a box distance of 40 pixels per 27.8 Åwhere the paper describes this · verbatim
The data points are available as Source data.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- ValidationNo validation of the AI is described.
- Version of SPHIRE_crYOLOWhich version of the model was used is not stated.
About this article
Record aix-00029, version 2, checked by a person on 2026-10-07. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY-4.0; quotations are at most 25 words. How we work · Report an error