Iterative Refinement for Structural Optimization of De Novo Protein Sequences

Iterative refinement of de novo protein sequences reduced mean RMSD from 8.097 Å to 4.570 Å on a matched sample of 224 proteins, improving predicted structures.

• Why it matters: Predicted de novo protein structures can retain substantial structural error after initial sequence generation, limiting accuracy for downstream design and analysis.
• What they did: Starting from structure-aware GNN predictions, the method iteratively substitutes residues, folds each candidate with ESMFold, and accepts changes meeting global TM-score/GDT-TS/RMSD improvement plus a local displacement threshold.
• The result: RMSD decreased 43.6% versus the initial prediction, with additional reductions of 46.2% relative to ProteinMPNN and 27.2% relative to native ESMFolded sequences (paired t-test, p < 0.001).