structural-biology/ai produced the result/ · v2
Deep learning reads cryo-EM maps alongside AlphaFold3 models to build protein structures
Researchers built MICA, a method that feeds a cryo-electron microscopy density map and AlphaFold3's predicted chain structures into one trained network, which labels each point in the map before the atomic model is assembled.
spectrum · one line per step, placed by what the step does · bright lines used AI
Multimodal deep learning integration of cryo-EM and AlphaFold3 for high-accuracy protein structure determination
doi:10.1101/2025.07.03.663071 · record aix-00168 v2 · checked 2026-10-09
- AI was for
- Structure determination, Detection, Classification
- Model family
- Convolutional neural network
- Checked by
- Benchmark80 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Proteins are long chains of amino acids that fold into particular three-dimensional shapes, and the shape largely decides what the protein does. One way to see that shape is cryo-electron microscopy, or cryo-EM, which freezes many copies of a protein and combines images of them into a three-dimensional map of density: a grid showing where matter is more or less concentrated. The map is not a structure, though. Someone still has to decide which blob of density is which atom, which amino acid sits where, and how the chain threads through the whole thing. Where the map is blurred or broken, that reading becomes guesswork.
A second source of information now exists. Prediction programmes such as AlphaFold3 can propose a protein's shape from its amino acid sequence alone, without any microscope. The two sources have different weaknesses: the map is experimental but patchy, the prediction is complete but not tied to this particular specimen. The researchers set out to use both at once, and to produce finished atomic models from cryo-EM maps automatically, comparing the results against structures already deposited in the Protein Data Bank and against two existing model-building methods.
Where AI came in
AlphaFold3 was run as it comes, off the shelf, to predict a structure for each protein chain from its sequence. Those predictions were cut into domains, fitted into the map's grid and turned into a numerical encoding. The researchers then trained their own network from scratch, a convolutional one of the kind used for image recognition but working on three-dimensional volumes. It takes the map and the encoded predictions through separate pathways, fuses them, and then answers three questions at every point in the grid: is this a backbone atom, is this a central carbon of an amino acid, and which of the twenty amino acids is it.
That per-point labelling is the part standing in for a person's reading of the map, which the record describes as displacing expert judgement. The steps after it are conventional software rather than learning: points above a probability threshold are grouped into candidate atoms, those candidates are traced into a chain against the known sequence, gaps are filled using the AlphaFold3 structures, side chains are added and the whole model is refined against the density. The record notes that the reported structures come from this combination, with no non-AI route to them described.
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
MICA is a deep learning method that builds atomic protein structures from cryo-EM density maps by taking the map and AlphaFold3-predicted chain structures together as input to a multi-task convolutional encoder-decoder with a feature pyramid network, which predicts backbone atoms, Cα atoms and amino acid types per voxel. These predictions are traced into a backbone model, gap-filled using the AlphaFold3 structures, converted to a full-atom model and refined against the map. On two test sets of 80 density maps each, MICA reported average TM-scores of 0.92 and 0.88, higher than EModelX(+AF) on all six evaluation metrics and higher than ModelAngelo on TM-score, Cα match, Cα quality score and aligned Cα length, while ModelAngelo reported higher sequence match on both datasets. On 12 EMDB maps released after 1 January 2025, the average TM-score was 0.93 and 9 of the 12 models had TM-score above 0.9.
How AI was used
AlphaFold3 was run off-the-shelf to predict a structure for each protein chain from its sequence. Each cryo-EM density map was resampled to 1 Å voxels and normalised, and the AF3 chain structures were split into domains with Merizo, docked into the map with phenix.dock_in_map, and encoded as a 24-channel binary volume covering backbone N, Cα, C, O atoms and the 20 amino acid types; maps, encodings and voxel-wise ground truth masks derived from PDB structures were partitioned into 64-voxel windows. A network trained from scratch processes the two modalities through separate multi-scale convolution and attention pathways, fuses them, passes them through three encoder blocks with residual dense and dual attention modules and a feature pyramid network, and uses three cascaded task-specific decoders to predict backbone atom, Cα atom and amino acid type labels; it was trained with a weighted cross-entropy loss whose task weights were shifted at epoch 25, with gradient clipping, dropout and on-the-fly augmentation. At inference the probability grids are stitched back to map dimensions, Cα voxels above a 0.3 probability threshold are grouped by DBSCAN with non-maximum suppression and coordinate refinement, and the resulting candidates are traced into a backbone model using EModelX(+AF)'s sequence-alignment and gap-filling protocol with the AF3 structures, then converted to full atoms with PULCHRA and refined with phenix.real_space_refine. ModelAngelo and EModelX(+AF) were run as comparison methods.
The shape of the work
Structural · the record, drawn
no AI
Assemble training, validation and test map sets
Obtaining raw data, whether by measurement, download or retrieval.
we collected 550 density maps released by April 2023 that exhibit high correlation with the true structures deposited in the Protein Data Bankwhere the paper describes this · verbatim
AI
Predict per-chain structures with AlphaFold3
Running a trained model over new data to predict, classify or score.
AlphaFold3 was used to predict the structure for each chain of the protein of each cryo-EM density map from its sequence.where the paper describes this · verbatim
no AI
Standardise maps, dock and encode AF3 structures, build label masks
Cleaning, filtering, normalising or labelling data already obtained.
we resampled each density map with a constant voxel spacing of 1 Å, ensuring consistency across all the mapswhere the paper describes this · verbatim
AI
Train multi-task encoder-decoder with feature pyramid network
Fitting model parameters, including fine-tuning an existing model.
The multimodal deep learning model of MICA was trained to predict backbone atoms, Cα atoms, and amino acid types simultaneously using weighted cross-entropy losswhere the paper describes this · verbatim
AI
Predict backbone atoms, Cα atoms and amino acid types
Running a trained model over new data to predict, classify or score. The AI stood in for expert judgement.
MICA predicts the positions of backbone atoms, Cα atoms and amino acid types of Cα atoms from the inputwhere the paper describes this · verbatim
no AI
Cluster and refine predicted Cα candidates
Cleaning, filtering, normalising or labelling data already obtained.
a DBSCAN clustering strategy is utilized to group predicted Cα voxels whose Cα probability exceeds a threshold of 0.3 into clusterswhere the paper describes this · verbatim
no AI
Trace backbone, fill gaps with AF3 structures, build and refine full-atom model
Producing candidate objects that did not previously exist.
EModelX(+AF)’s backbone tracing protocol is used to build a Cα backbone modelwhere the paper describes this · verbatim
no AI
Score models against PDB structures and baseline methods
Testing outputs against ground truth.
The atomic models built by the three methods from density maps were compared with the corresponding ground truth structures from the Protein Data Bank (PDB)where 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 paper's result is the atomic structural models themselves, which are produced by the trained multi-task network plus AlphaFold3 structures; no non-AI route to the reported models is described
Multiple 3D convolutional filters with kernel sizes of 3, 5, 7, and 9 in the block generate multi-scale featureswhere the paper describes this · verbatim
This filtering resulted in 80 density maps left in each dataset to compare the three methods.where the paper describes this · verbatim
The scripts, programs and instructions for downloading files, processing data and running the MICA are found inwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of MICAWhich version of the model was used is not stated.
- Version of AlphaFold3Which version of the model was used is not stated.
- Version of ModelAngeloWhich version of the model was used is not stated.
- Version of EModelX(+AF)Which version of the model was used is not stated.
- Version of MerizoWhich version of the model was used is not stated.
- What step 2 replacedThe paper gives no basis for what the AI stood in for.
- What step 4 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00168, version 2, checked by a person on 2026-10-09. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error