RESEARCH · RESEARCH · #1248
Mirror-Score: calibrated, inference-only scoring and public benchmark for D-peptide/L-protein design (arXiv:2609.36057v1)
This paper introduces Mirror-Score, a calibrated inference-only scoring framework and a public benchmark (31 crystal complexes, 18 with known affinities) for heterochiral D-peptide/L-protein complexes, with code and data at the linked GitHub. The authors show that raw ProteinMPNN negative log-likelihood (NLL) is an unreliable affinity ranker across families (pooled Spearman rho = 0.19, with sign reversals between targets) and instead highlight Boltz-2 mirror-space cofolding confidence and interface pLDDT as more informative for family-matched ranking (e.g., structure-level leave-one-out Spearman rho = 0.90 on the viral-entry family), while noting limited cross-family transfer at current sample sizes.
KEY POINTS
- This paper introduces Mirror-Score, a calibrated inference-only scoring framework and a public benchmark (31 crystal complexes, 18 with known affinities) for heterochiral D-peptide/L-protein complexes, with code and data at the linked GitHub.
- The authors show that raw ProteinMPNN negative log-likelihood (NLL) is an unreliable affinity ranker across families (pooled Spearman rho = 0.19, with sign reversals between targets) and instead highlight Boltz-2 mirror-space cofolding confidence and interface pLDDT as more informative for family-matched ranking (e.g., structure-level leave-one-out Spearman rho = 0.90 on the viral-entry family), while noting limited cross-family transfer at current sample sizes.
- This matters because it empirically demonstrates that sequence-compatibility NLL from ProteinMPNN is not a reliable cross-family affinity ranker and provides a calibrated, inference-only framework and benchmark to improve practical D-peptide design and evaluation.
WHY IT MATTERS
This matters because it empirically demonstrates that sequence-compatibility NLL from ProteinMPNN is not a reliable cross-family affinity ranker and provides a calibrated, inference-only framework and benchmark to improve practical D-peptide design and evaluation.