An Automated Computational Factory for High-Confidence De Novo Protein Design: Multi-Temperature Sequence Design with Physics-Based Validation

Abstract

We present an automated computational design factory integrating RFdiffusion backbone generation, ProteinMPNN multi-temperature sequence design, ColabFold (AlphaFold2) structure prediction, and OpenMM molecular-dynamics physics validation. Over a sustained production run we generated a library of 2,422 unique de novo proteins (median pLDDT 91.1, all pTM >= 0.70, median solubility 92.6). A rebalanced composite score (0.35*pLDDT + 0.35*100*pTM + 0.30*solubility) reaches max 95.2 (median 88.5). A stall-detection termination criterion reduces redundant re-entry by ~70%. The companion dataset is available at DOI 10.5281/zenodo.22924852.