Iterative Refinement for Structural Optimization of De Novo Protein Sequences

Abstract

In this paper, we introduce an iterative refinement procedure that substitutes individual amino acid residues in a predicted protein sequence to minimize structural error; substitutions are guided by a weighted composite score of global evaluation metrics including the TM-score, GDT-TS, and RMSD. Starting from an initial sequence prediction produced by a structure-aware graph neural network (GNN) [5], the refinement procedure evaluates candidate amino acid substitutions at each position by folding the modified sequence with ESMFold [9] and accepting substitutions that satisfy both a global structural improvement criterion and a local displacement threshold. Evaluated on a matched sample of 224 proteins, this set of methods reduced mean RMSD from 8.097 [A] to 4.570 [A] relative to the model's initial prediction, a 43.6% reduction, with corresponding reductions of 46.2% relative to ProteinMPNN and 27.2% relative to native ESMFolded sequences; these differences were statistically significant by paired t-test (p < 0.001). Our results show that post-prediction iterative refinement can significantly improve the structural accuracy of de novo protein sequences.