Scaffold-Constrained Generative Design Overcomes Failure Modes in Computational Protein Engineering.

Applying unconstrained generative protein models to fixed structural scaffolds can produce systematic design artifacts, including glycine enrichment, grammar violations, and discordant complex predictions. We developed DARPinMPNN, a scaffold-constrained computational framework that restricts production sequence generation to permitted DARPin Greek Grammar residue pools. A state-aware chimeric multiple sequence alignment strategy supported AlphaFold2-Multimer (AF2) screening, followed by complementary AlphaFold 3 (AF3) evaluation. Of 15,000 generated variants targeting mesothelin, VEGF-A, and HER2, 149 library designs and five separately analyzed controls had deposited 25-model AF3 results. Using the unrounded median of 25 corrected target-DARPin Ranking Scores, eight library designs passed the threshold of 0.70, all targeting mesothelin, corresponding to 0.053% of the generated library. The highest-ranked design had median corrected RS 0.794; the G3 positive control passed separately with median corrected RS 0.866. AF3 median chain-pair ipTM was lower than AF2 ipTM for 148 of 149 library designs. These results demonstrate target-dependent computational prioritization and cross-predictor discordance, not experimental binding or false-positive rates. The candidates require biophysical and cellular validation.