AlphaFold

The Evoformer is AlphaFold's central neural network block that processes the Multiple Sequence Alignment (MSA) and pairwise residue interactions simultaneously. It uses a novel mechanism where information flows bidirectionally between the MSA representation and the pair representation through axial attention—applying attention across rows (sequences) and columns (residue positions) independently. This allows the model to reason about evolutionary covariation and spatial proximity jointly, capturing both co-evolutionary signals and geometric constraints in a single unified architecture.

The Structure Module translates abstract pair representations into explicit 3D atomic coordinates using an equivariant transformer architecture. Unlike traditional deep learning layers that operate on arbitrary vectors, this module respects the symmetries of 3D space—rotating or translating the input produces an identically transformed output. It predicts a residue gas representation where each amino acid is treated as a rigid body with a rotation matrix and translation vector, iteratively refining the backbone geometry through 8 cycles of Invariant Point Attention (IPA). This design ensures the predicted structure is physically plausible without requiring post-prediction energy minimization.

AlphaFold employs a recycling mechanism where the model's initial predictions are fed back as input for multiple passes, typically 3–4 iterations. At each cycle, the predicted pairwise distances and orientations refine the MSA and pair representations, allowing the model to progressively resolve structural ambiguities. This technique effectively performs in-silico self-distillation—the model learns from its own intermediate outputs to correct errors in side-chain packing and domain orientation. Recycling is particularly critical for challenging targets with shallow MSAs or multi-domain architectures where initial predictions may contain topological errors.

AlphaFold outputs two critical per-residue confidence metrics that guide experimental interpretation. The predicted Local Distance Difference Test (pLDDT) scores each residue from 0–100, estimating how well the local atomic environment matches the true structure—residues with pLDDT > 90 are considered high-confidence and suitable for detailed analysis. The Predicted Aligned Error (PAE) matrix quantifies the expected positional error between any pair of residues when the predicted and true structures are optimally aligned. Low PAE values between domains indicate confident relative positioning, while high PAE suggests flexible linkers or domain uncertainty. These metrics are essential for cryo-EM model docking and assessing domain-level reliability.

AlphaFold-Multimer extends the original architecture to predict the structures of protein-protein complexes by modifying the input featurization and training on the PDB's biological assembly data. Key innovations include chain-relative position encoding to distinguish intra-chain from inter-chain residue pairs, and a cross-chain MSA representation that captures co-evolutionary signals across interaction interfaces. The model achieves high accuracy for obligate complexes and antibody-antigen pairs, though performance degrades for transient interactions or complexes with significant conformational changes upon binding. It outputs a unified structure with all chains simultaneously, avoiding the need for computational docking of individually predicted monomers.

AlphaFold was trained on the Protein Data Bank (PDB) using a supervised learning paradigm with approximately 365,000 experimentally determined structures. The training pipeline includes extensive data augmentation through genetic database search against BFD, UniRef90, and MGnify clusters to construct rich MSAs. A critical innovation is the self-distillation process where the model generates predictions on UniClust30 sequences, and high-confidence predictions are added to the training set—effectively expanding the training data to millions of structures. The loss function combines the Frame Aligned Point Error (FAPE) for backbone accuracy with auxiliary losses on distogram, torsion angles, and masked MSA reconstruction, ensuring the model learns both local geometry and global topology.

AlphaFold predictions serve as superior search models for molecular replacement in cryo-EM, even at low resolutions where traditional homology models fail.

• Enables phasing of maps at 4-5 Å resolution where backbone tracing is ambiguous
• Outperforms Rosetta and homology models in map-to-model correlation
• Particularly effective for novel folds with no known structural homologs
• Reduces the resolution barrier for de novo structure determination

ModelAngelo integrates AlphaFold predictions with graph neural networks to automate atomic model building directly into cryo-EM density maps.

• Uses AlphaFold's confidence metrics (pLDDT) to guide sequence assignment
• Combines GNN-based backbone tracing with predicted residue probabilities
• Achieves near-complete models without manual intervention for maps better than 3.5 Å
• Dramatically reduces model building time from weeks to hours

AlphaFold structures are used as starting coordinates for molecular dynamics flexible fitting (MDFF) to capture conformational changes induced by the cryo-EM environment.

• MDFF applies density-derived forces to flexibly fit AlphaFold models
• Resolves domain movements and loop rearrangements not captured in the prediction
• Validates AlphaFold's accuracy by measuring map-model cross-correlation
• Identifies regions where the solution state differs from the predicted conformation

AlphaFold predictions provide multiple conformational states for initializing heterogeneous refinement, helping to resolve compositional and conformational heterogeneity.

• AlphaFold's multimer predictions generate plausible assembly states
• Used as 3D references for maximum-likelihood classification in RELION and cryoSPARC
• Enables separation of apo vs. holo states and ligand-bound conformations
• Validates the biological relevance of predicted alternative conformations

AlphaFold models serve as ground truth references for validating cryo-EM maps and identifying regions of genuine structural novelty.

• FSC comparison between AlphaFold model and experimental map quantifies agreement
• Q-score analysis identifies well-resolved vs. disordered regions
• Highlights novel structural features absent from the prediction (e.g., ligands, cofactors)
• Guides manual rebuilding efforts to regions where the prediction deviates from density

AlphaFold predictions generate high-resolution templates for subtomogram averaging in cryo-ET, where individual particles are extracted from crowded cellular environments.

• Provides initial references for aligning low-signal subtomograms
• Enables in situ structure determination of complexes within native cellular context
• Bridges the gap between purified systems and cellular structural biology
• Validated against EMPIAR subtomogram averaging benchmarks with high correlation

The foundational input to AlphaFold's architecture. MSAs are computed by searching sequence databases for evolutionarily related proteins, revealing co-evolutionary couplings between residue pairs. These couplings encode spatial proximity constraints—if two residues mutate in tandem across species, they are likely close in 3D space. AlphaFold's Evoformer block processes the MSA to extract these pairwise relationships, which are then used to predict inter-residue distances and orientations. The depth and quality of the MSA directly correlate with prediction accuracy; shallow MSAs for orphan proteins or antibodies remain a key limitation.

An inverse folding neural network developed by the Baker Lab that complements AlphaFold's sequence-to-structure prediction. Given a protein backbone structure, ProteinMPNN designs amino acid sequences predicted to fold into that shape. This is critical for cryo-EM model validation: after building an atomic model into a density map, ProteinMPNN can verify whether the local backbone geometry supports the assigned sequence. It is also used to redesign protein interfaces for increased stability or novel binding functions, effectively closing the loop between structure prediction and sequence design.

A class of neural architectures that guarantee outputs transform predictably under 3D rotations and translations of the input—a property called SE(3)-equivariance. AlphaFold's Structure Module uses an invariant point attention (IPA) mechanism that respects these symmetries, ensuring the predicted structure is independent of the input coordinate frame. Unlike traditional CNNs that require data augmentation to learn rotational invariance, equivariant networks bake this physical constraint into their mathematical structure, dramatically improving sample efficiency and prediction consistency for molecular systems.

An automated atomic model building program that uses a graph neural network to trace the protein backbone and assign amino acid sequences directly into cryo-EM density maps. ModelAngelo integrates AlphaFold predictions as a strong prior, using the predicted structure to guide chain tracing through noisy or ambiguous density regions. This synergy between prediction and experimental data accelerates structure determination from weeks to hours. The GNN operates on a graph representation of the density map, classifying nodes as backbone atoms and edges as peptide bonds.

A method that uses molecular dynamics simulation to flexibly fit an atomic model into a cryo-EM density map by applying forces derived from the map's potential. AlphaFold predictions serve as the initial model, which is then driven into the experimental density through simulated annealing. The MD engine applies physical force fields to maintain stereochemical plausibility while external forces proportional to the density gradient pull atoms into high-density regions. This resolves clashes and refines side-chain rotamers that AlphaFold may have predicted with lower confidence.

A deep generative model using a variational autoencoder to reconstruct continuous conformational heterogeneity from cryo-EM images. While AlphaFold predicts a single static structure, CryoDRGN learns a latent space of structural states directly from 2D particle images. The encoder maps each particle image to a latent coordinate, and the decoder generates a 3D density map from that coordinate. This reveals the full conformational landscape—motions, intermediates, and rare states—that a single AlphaFold prediction cannot capture, making the two tools complementary for understanding protein dynamics.