D‑peptides combine protease resistance with high target specificity, yet computational design of D‑peptide binders is still immature. Mirror‑Peptidizer introduced an in‑silico mirror‑image screening pipeline that uses target reflection, backbone generation and ProteinMPNN sequence design, but it ranked candidates directly by ProteinMPNN negative log‑likelihood (NLL) without validation against measured affinities; only 4 of 9 MDM2 designs showed detectable binding.
To address this gap we present Mirror‑Score, a calibrated inference‑only scoring framework for heterochiral D‑peptide/L‑protein complexes, together with a public benchmark of 31 crystal complexes spanning four target families, 18 of which have literature‑verified affinities. Raw ProteinMPNN NLL proved to be an unreliable affinity ranker: pooled Spearman correlation with affinity is only 0.19, and the sign flips between families (MDM2/CHIP +0.62 vs gp41 –0.70).
We therefore evaluated Boltz‑2 mirror‑space co‑folding confidence. For the complete viral‑entry family (7 structures representing 3 peptides), interface predicted local distance difference test (pLDDT) achieved a structure‑level leave‑one‑out Spearman ρ of 0.90 (p = 0.006) and correctly ordered all three peptides by affinity, whereas NLL yielded ρ = 0.18. Because only three independent chemotypes are represented, this result reflects directional consistency rather than a statistically validated predictor.
Cross‑family calibration does not transfer at current sample sizes, supporting family‑matched calibration as the practical deployment mode. We also outline a prospective design protocol for the antimicrobial‑resistance targets LasR and LecB from Pseudomonas aeruginosa, providing mirrored structures, ligand‑derived hotspot maps, diffusion‑model‑ready inputs and Mirror‑Score rankings. All code, benchmark data, structures and analysis scripts are openly available at https://github.com/Jiadalee/Mirror-Score.
Review: Mirror‑Score replaces raw NLL with calibrated pLDDT, delivering reliable affinity ranking under limited data conditions and establishing a reproducible benchmark for heterochiral peptide design. Expanding family‑level datasets should further improve cross‑family predictive power.