structural-biology/ai produced the result/Nature 2023 · v2
Cryo-tomography maps the myosin filament inside relaxed mouse heart muscle
Researchers imaged relaxed mouse cardiac muscle by cryo-electron tomography and built an atomic model of its thick filament. Learned software traced the filaments, cleaned up images, and predicted the shapes of the proteins fitted into the density.
spectrum · one line per step, placed by what the step does · bright lines used AI
Structure of the native myosin filament in the relaxed cardiac sarcomere
Nature, 2023
doi:10.1038/s41586-023-06690-5 · record aix-00005 v2 · checked 2026-10-07
- AI was for
- Structure determination, Detection, Denoising
- Model family
- Transformer, Convolutional neural network
- Checked by
- Experimental
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Heart muscle contracts because two sets of protein filaments slide past each other. The thin filaments are built around actin; the thick filaments are built from myosin, the motor protein whose heads pull on actin. These filaments are packed into repeating units called sarcomeres, along with a giant elastic protein, titin, and a regulatory protein called cardiac myosin-binding protein C. Working out how all these pieces are arranged is hard. The arrangement only makes sense inside intact muscle, where the filaments sit in their natural lattice, and the parts are long, flexible and crowded together. Images of frozen tissue are also very noisy, so individual molecules are difficult to see.
The researchers set out to determine the structure of the thick filament in place, in relaxed muscle, rather than in purified preparations. They used muscle fibres from the left ventricle of the mouse heart, chemically held in a relaxed state, thinned with a focused ion beam and imaged by cryo-electron tomography, which records tilted views of a frozen specimen and reconstructs a three-dimensional volume. They then averaged many copies of the same filament segments and assigned which protein occupied which part of the resulting density.
Where AI came in
Three pieces of learned software sat inside an otherwise conventional imaging pipeline. A trained picker, SPHIRE-crYOLO, was run over the reconstructed volumes to find and trace both the thick and the thin filaments, supplying the coordinates from which segments were cut out for averaging; this took the place of picking them out by hand. A denoising network, cryo-CARE, was used to clean up two representative tomograms so that the components could be segmented and the flexible links of myosin-binding protein C traced by hand.
AlphaFold2, a protein structure prediction system, was given amino acid sequences in overlapping pieces: the myosin tail, the tail end of myosin-binding protein C, and a stretch of titin. Its predicted models, with published experimental structures, were fitted into the measured density. The record notes that the atomic-level claims depend on these models, which were used to set the position and register of the titin and myosin segments in place of determining those shapes experimentally.
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
Cryo-electron tomography of relaxed, demembranated mouse cardiac myofibrils was used to determine the in situ structure of the thick filament from the M band through the P zone to the C zone. A learned particle picker traced the thick and thin filaments in the tomograms, subtomogram averaging produced a thin filament map at 8.2 Å resolution and eight thick filament segment maps at 19.3 Å to 23.6 Å, and AlphaFold2 predictions of myosin tails, cMyBP-C and titin were fitted into the density to assign the components. The resulting model describes three distinct myosin crown arrangements, three pairs of titin-α and titin-β chains in which titin-β stops at crown A1, and cMyBP-C C-terminal domains that contact myosin tails and the free heads of crowns 1 and 3 while the N-terminal region links to thin filaments. An affinity-purified antibody and stimulated emission depletion microscopy placed the titin kinase domain 78 nm ± 8 nm from the M1 line.
How AI was used
Three learned models were used inside an otherwise conventional cryo-ET pipeline. After motion correction, tilt series alignment and tomogram reconstruction in Warp and IMOD, SPHIRE-crYOLO was run on binned, low-pass-filtered tomograms to pick and trace both thick and thin filaments, giving the coordinates from which subtomograms were resampled and extracted; the subsequent 2D classification, 3D classification and helical refinement were carried out in ISAC and RELION. AlphaFold2 was then run on amino acid sequences submitted in overlapping segments — the MYH7 myosin tail, the C-terminal region of MYBPC3, and titin from domain A101 to m3 — to produce atomic models that, together with published experimental structures, were placed into the reconstructions by rigid body fitting and refined by molecular dynamics flexible fitting in Namdinator; the predictions were also used to set the register and position of titin domains and to identify a predicted kink in the myosin tail. Separately, cryo-CARE was used to denoise two representative tomograms before back-plotting the refined structures as binary masks, pseudo-segmenting the volumes in Dragonfly, and manually tracing the flexible cMyBP-C links.
The shape of the work
Structural · the record, drawn
no AI
Prepare relaxed myofibrils and collect tilt series
Obtaining raw data, whether by measurement, download or retrieval.
All images and a total of 89 tomograms were acquired using SerialEM.where the paper describes this · verbatim
no AI
Align tilt series and reconstruct tomograms
Cleaning, filtering, normalising or labelling data already obtained.
Motion correction and contrast transfer function estimation were carried out in Warpwhere the paper describes this · verbatim
AI
Pick and trace thick and thin filaments
Cleaning, filtering, normalising or labelling data already obtained. The AI stood in for manual curation.
we used SPHIRE-crYOLO to pick and trace both the thick and the thin filamentswhere the paper describes this · verbatim
no AI
Classify subtomograms and average filament segments
Cleaning, filtering, normalising or labelling data already obtained.
resulting in 37,118 high-quality particles that were re-extracted as subtomograms.where the paper describes this · verbatim
AI
Predict atomic structures of filament components
Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.
the tails were predicted in AlphaFold2 using the amino acid sequence of MYH7 from Mus musculuswhere the paper describes this · verbatim
no AI
Fit models into density and assign components
Extracting understanding from model behaviour.
The models were initially built in the map using rigid body fittingwhere the paper describes this · verbatim
AI
Denoise representative tomograms
Cleaning, filtering, normalising or labelling data already obtained. The AI stood in for conventional algorithm.
denoised them using cryo-CAREwhere the paper describes this · verbatim
no AI
Back-plot, segment and trace cMyBP-C links
Extracting understanding from model behaviour.
After clearly identifying and segmenting 76 cMyBP-C links in our tomograms, we measured the angle that the link formedwhere 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 reported architecture rests on AlphaFold2 models to assign densities and fix the register of titin, myosin tails and cMyBP-C, and on crYOLO to pick and trace the filaments that were averaged; the atomic-level claims do not stand without them
we generated a new affinity-purified TK domain antibody and used it to ascertain the TK domain position by super-resolution microscopywhere 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.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of AlphaFold2Which version of the model was used is not stated.
- Version of SPHIRE-crYOLOWhich version of the model was used is not stated.
- Version of cryo-CAREWhich version of the model was used is not stated.
About this article
Record aix-00005, 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