AlphaFold2

A novel transformer-based module that processes Multiple Sequence Alignments (MSAs) and pairwise residue representations simultaneously. Unlike standard transformers, the Evoformer uses axial attention to handle the high-dimensional MSA tensor efficiently. It infers co-evolutionary couplings and spatial proximity constraints directly from sequence data, producing a refined residue pair representation that encodes the predicted distance and orientation between every pair of amino acids. This representation serves as the geometric blueprint for the subsequent structure module.

The final stage of the network that translates the abstract pair representation into explicit 3D atomic coordinates. Its core innovation is the Invariant Point Attention (IPA) mechanism, which performs attention over the 3D spatial relationships between residues. IPA guarantees that the model's predictions are invariant to global rotation and translation of the entire protein, a critical physical constraint. The module iteratively updates a local coordinate frame for each residue, predicting backbone torsion angles and side-chain positions to construct the final all-atom structure.

A mechanism where the model's initial predicted structure and refined representations are fed back as input for multiple passes, typically three. Each recycling iteration allows the network to resolve internal inconsistencies and progressively improve accuracy. The predicted 3D coordinates from one pass inform the attention mechanisms in the next, creating a feedback loop that mimics an energy minimization process. This iterative refinement is crucial for achieving sub-angstrom accuracy on challenging targets and for correctly packing complex domain architectures.

The entire AlphaFold2 system is trained as a single, end-to-end differentiable model. The loss function directly compares the predicted 3D coordinates to the experimental ground truth using the Frame Aligned Point Error (FAPE) loss. FAPE measures coordinate error after optimal alignment, providing a more robust training signal than traditional RMSD. This holistic training allows gradients to flow from the final 3D structure back through the Structure Module and into the Evoformer, jointly optimizing all parameters for the ultimate goal of atomic-accuracy prediction.

AlphaFold2 outputs two essential per-residue and pairwise confidence metrics that are integral to its architecture, not post-hoc calculations. The predicted Local Distance Difference Test (pLDDT) scores each residue from 0-100, indicating local backbone reliability. The Predicted Aligned Error (PAE) matrix estimates the expected positional error between any two residues, revealing domain packing accuracy. These self-assessment tools allow researchers to identify well-predicted structured domains and distinguish them from flexible, low-confidence intrinsically disordered regions.

In addition to MSAs, the architecture optionally ingests structural templates from the Protein Data Bank (PDB) identified through sequence homology. These templates are processed by a dedicated embedding network that extracts pairwise distance and orientation features from the experimental structures. This information is integrated into the Evoformer's pair representation, providing a direct geometric prior that is particularly valuable for modeling proteins with known structural homologs, improving accuracy on conserved folds.

A foundational input feature for AlphaFold2 that aligns evolutionarily related protein sequences to identify conserved regions and co-evolving residues. MSAs provide the model with rich evolutionary constraints that implicitly encode spatial proximity information. AlphaFold2 processes MSAs through its Evoformer block, which performs row-wise and column-wise attention to extract residue-residue interaction patterns. The quality and depth of the MSA directly impacts prediction accuracy—shallow alignments for orphan proteins remain a key limitation.

The core architectural innovation of AlphaFold2's Structure Module that performs attention over 3D spatial relationships while maintaining invariance to global rotation and translation. IPA operates on a residue gas representation where each residue is defined by a 3D coordinate frame. The attention mechanism computes pairwise biases based on Euclidean distances and relative orientations between frames, ensuring that predicted structures are physically consistent regardless of the coordinate system. This design eliminates the need for data augmentation over rotations.

An iterative refinement mechanism where AlphaFold2's initial predicted structure is fed back as input for multiple passes, progressively improving accuracy. During recycling, the model's output pairwise representations and backbone coordinates from the previous iteration are concatenated with the input features. This allows the network to iteratively correct errors in side-chain packing and backbone geometry. Typically, 3 recycling iterations are used during inference, with diminishing returns beyond this point. Recycling is particularly effective for improving predictions of multi-domain proteins.

A per-residue confidence metric output by AlphaFold2 that estimates the local accuracy of the predicted structure on a scale from 0 to 100. pLDDT is derived from a self-consistency check where the model predicts the lDDT-Cα score it would receive if compared to an experimental reference. Residues with pLDDT > 90 are modeled with high accuracy and suitable for detailed structural analysis, while regions with pLDDT < 50 may indicate intrinsically disordered regions (IDRs) and should be interpreted with caution. This metric is crucial for downstream applications like drug docking.

A pairwise confidence metric that estimates the expected positional error between any two residues in a predicted structure. The PAE matrix is an N×N heatmap where each cell (i, j) indicates the expected error in residue i's position if the predicted and true structures were aligned on residue j. Low PAE values between domains indicate confident relative domain orientations, while high PAE suggests flexible linkers or uncertain packing. PAE is essential for assessing domain-domain orientation accuracy and identifying rigid structural modules within flexible proteins.

A biennial community-wide blind experiment that rigorously benchmarks computational protein structure prediction methods. CASP14 in 2020 marked the historic breakthrough where AlphaFold2 achieved median GDT_TS scores competitive with experimental methods, effectively solving the 50-year-old protein folding problem. Predictors receive amino acid sequences for proteins whose structures have been experimentally determined but not yet publicly released, ensuring a true blind test. CASP evaluates predictions using multiple metrics including GDT_TS, lDDT, and TM-score.