MICA

Cryo-electron microscopy now routinely resolves large protein complexes, but converting a density map into an accurate atomic model remains one of the slowest steps in the pipeline. Automated builders that trace chains directly from density degrade sharply where the map is noisy, locally low-resolution, or incomplete. Structure predictors solve a complementary problem: they produce complete, chemically sensible models from sequence alone, but those models are untethered from the experimental data and can be wrong in exactly the regions a map would constrain.

MICA, developed by Rajan Gyawali, Ashwin Dhakal, and Jianlin Cheng at the University of Missouri, combines both signals. It fuses cryo-EM density maps with AlphaFold 3 predictions at two points: as joint input to a network that predicts backbone atoms, Cα positions, and residue identities, and again at the output stage, where predicted structures complete and refine the traced backbone into a final atomic model. The work appeared on bioRxiv in July 2025 and in Communications Chemistry that October.

Input-level fusion is what separates MICA from earlier hybrid approaches, which use predicted structures mainly as templates fitted to density after tracing. Seeing density and prediction together lets the network arbitrate between them voxel by voxel rather than committing to one source and patching with the other.

#Key Features

• Dual-level fusion: AlphaFold3 predictions enter both as network input features aligned to the density grid and as a source of structural completion after backbone tracing, rather than only as post-hoc templates.
• Multi-task prediction heads: One encoder-decoder predicts backbone atoms, Cα atoms, and amino acid type at once, so chain tracing and sequence registration share a representation.
• Learned distrust of the prediction: A feature-gate module downweights unreliable predicted regions, and 40% of training samples withhold the AlphaFold3 channels entirely so the network cannot become dependent on them.
• Resilience to poor predictions: When the input AlphaFold3 model scores below 0.5 TM-score, MICA still reaches 0.77; with high-quality predictions (TM-score ≥ 0.8) it reaches 0.96.
• End-to-end automation: A FASTA sequence and a density map are the only inputs; domain parsing, docking, prediction, model building, and side-chain reconstruction run as one pipeline.

#Technical Details

Density maps are resampled to 1 Å voxel spacing, normalized against a median-based background estimate, and cut into 64³ voxel cubes with eight voxels of contextual padding. Docked AlphaFold3 structures are encoded as a 24-channel binary grid covering Cα, N, C, and O atoms plus the 20 amino acid types. A multi-scale convolution block with kernel sizes of 3, 5, 7, and 9 expands the input to 128 channels, feeding three encoder stages built from residual dense blocks with local and global attention and squeeze-and-excitation gating; a feature pyramid network with learnable per-scale weights aggregates the multi-resolution features, and three decoders emit backbone (4 channels), Cα (4 channels), and amino acid (21 channels) predictions.

Training used 550 density maps from Cryo2StructData released through April 2023, split 440 for training and 110 for validation, with test proteins filtered below 25% sequence identity to the training set. On an 80-map Cryo2StructData test set, MICA reached a mean TM-score of 0.92 with 93.71% Cα matching, against 0.89 for EModelX(+AF) and 0.75 for ModelAngelo; on a second 80-map benchmark spanning 1.52–3.99 Å and up to 11,109 residues, it scored 0.88 versus 0.85 and 0.57. On 12 maps released after the training cutoff, the average TM-score was 0.93. Runtime spans 10–30 minutes for 300–700 residues to over four hours above 2,500 residues on 24 CPU cores and one A100, with Phenix domain docking taking 85–97% of wall time.

#Applications

MICA targets the model-building stage of a standard cryo-EM workflow, taking a reconstructed map and the target sequence and returning an atomic model a structural biologist can inspect and correct rather than build from scratch. It is most useful for large multi-chain complexes, where manual tracing is slowest and predicted per-chain structures give the most leverage, and for maps with heterogeneous local resolution, where density-only tracing tends to break. The pipeline depends on external tools — Merizo for domain parsing, Phenix for docking, PULCHRA for side chains — so it fits laboratories already running a conventional cryo-EM software stack.

#Impact

MICA is a peer-reviewed demonstration that experimental density and predicted structure are more useful jointly than in sequence, and its released artifacts make that claim checkable: MIT-licensed code, trained weights, and the training, test, and results archives are public under CC BY 4.0. The authors are explicit about where it falls short. Accuracy on some large complexes is limited by chain and sequence registration errors when density is missing or noisy, performance is uneven beyond 4 Å, and the network predicts a Cα backbone, with all-atom detail added by post-processing rather than modeled directly. Depending on AlphaFold3 also inherits that model's access terms and failure modes.