Recycling

Recycling transforms AlphaFold2 from a single-pass predictor into a recurrent architecture. The output pair representation and single representation from one iteration become the input for the next, allowing the model to refine its own predictions without external intervention. This loop is unrolled for a fixed number of iterations (typically 3-4) during both training and inference.

During training, gradients flow through all recycling iterations, effectively creating a deep computational pathway that learns to correct systematic errors. Key aspects:

• The model learns to identify and fix stereochemical violations
• Early iterations focus on global topology, later iterations on local geometry
• Truncated backpropagation is used to manage memory constraints

The pair representation—an N×N matrix encoding relationships between all residue pairs—is the primary carrier of information across recycling iterations. It is updated by:

• Adding the output pair representation to the input via a residual connection
• No learned gating or weighting is applied; the addition is direct
• This allows the model to accumulate pairwise distance and orientation constraints incrementally

The single representation (MSA row-wise embeddings) is also recycled, but with a learned gating mechanism. The model computes a gated sum of the previous and current single representations, allowing it to selectively retain or overwrite per-residue information. This prevents the accumulation of noise while preserving useful structural signals.

In AlphaFold-Multimer and newer variants, recycling operates without re-querying the MSA. The initial MSA embedding is computed once, and subsequent iterations refine the structure purely from the pair and single representations. This decoupling:

• Reduces computational cost significantly
• Demonstrates that recycling learns a structure-conditioned refinement independent of evolutionary signal
• Enables faster inference for large complexes

Ablation studies show that removing recycling causes a significant drop in TM-score, particularly for:

• Hard targets with few homologous sequences
• Multi-domain proteins requiring correct inter-domain packing
• Side-chain accuracy and rotamer recovery Recycling provides the largest accuracy gain after the Evoformer itself, making it indispensable for atomic-accuracy predictions.

The core spatial reasoning mechanism that operates alongside recycling. IPA performs attention over the 3D coordinates of residues, ensuring that predictions remain invariant to global rotation and translation. It uses a novel architecture that represents each residue as a local reference frame, allowing the model to reason about pairwise spatial relationships without being confused by the protein's overall orientation in space.

The primary evolutionary input that recycling refines. MSAs align homologous protein sequences to identify co-evolutionary couplings between residue pairs. These couplings provide distance constraints that guide the initial structure prediction. The recycling mechanism iteratively re-interprets the MSA representation, allowing the model to extract increasingly subtle co-evolutionary signals that were not resolved in the first pass.

The per-residue confidence metric that validates recycling's effectiveness. pLDDT scores range from 0 to 100 and estimate the local accuracy of the prediction. Regions with low pLDDT scores often correspond to intrinsically disordered regions or domains that require more recycling iterations to converge. Monitoring pLDDT improvement across recycling passes provides a diagnostic for whether the model is genuinely refining the structure or oscillating.

The final refinement step applied after recycling concludes. While recycling operates on the neural network's internal representations, energy minimization uses a physics-based force field to resolve steric clashes and bond geometry violations in the output coordinates. This post-processing step corrects unphysical local geometries that the recycling process may have introduced, ensuring the final model is chemically plausible.

A limitation that recycling alone cannot fully address. Proteins are dynamic molecules that sample multiple structural states. Standard AlphaFold2 recycling typically converges to a single static structure, potentially missing alternative conformations relevant to function. Recent advances combine recycling with denoising diffusion models to generate diverse ensembles that capture the intrinsic flexibility of the protein.

The blind benchmarking experiment that validated recycling's contribution. CASP14 demonstrated that AlphaFold2's iterative refinement approach achieved median GDT_TS scores above 90 across all targets, effectively solving the single-domain protein folding problem. The recycling mechanism was identified as a critical component enabling the model to escape local minima and achieve atomic accuracy on challenging targets.