De novo Protein Generating Model via Structure-Aware Graph Neural Networks

A structure-aware graph neural network that optimizes 3D conformational accuracy via ESMFold achieves RMSD, TM-score, and GDT-TS statistically indistinguishable from ProteinMPNN on a matched evaluation sample.

• Why it matters: Sequence-recovery supervision may not be required for accurate protein structure prediction, and the field needs models that target structural precision for de novo protein generation.
• What they did: The model takes protein backbone coordinates as input and predicts amino acid sequences using a composite loss that heavily weights structural similarity, integrating differentiable structure prediction through ESMFold.
• The result: Despite substantially lower native sequence recovery and no direct sequence-recovery supervision, purely structural training yields competitive structural accuracy, suggesting multiple non-unique sequence solutions and enabling de novo generation on non-natural proteins.