A Benchmark for Heterochiral D-Peptide Design
On September 30, 2026, researchers released a paper (arXiv:2609.36057) introducing Mirror-Score, a standardized evaluation framework for heterochiral D-peptide/L-protein design. The team published a curated benchmark comprising 31 complex crystal structures across four target families, including 18 structures with validated experimental binding affinity data.
D-peptides possess inherent resistance to natural protease degradation and exhibit high target specificity, yet computational design tools for this molecular class remain constrained. Previously, the Mirror-Peptidizer pipeline integrated target mirroring with the ProteinMPNN model for sequence design. However, using raw ProteinMPNN negative log-likelihood (NLL) scores had never been benchmarked against measured binding affinities, leading to low success rates where only four of nine MDM2 designs bound experimentally.
Limitations of ProteinMPNN and Alternative Cofolding Metrics
Experimental results across the 31-complex benchmark revealed that raw ProteinMPNN NLL is not a valid proxy for binding affinity: - Weak Global Correlation: The overall Spearman correlation reached only 0.19 across the benchmark. - Opposing Trends Across Target Families: Correlations diverged drastically, yielding a positive correlation of +0.62 on MDM2/CHIP but reversing to an inverse correlation of -0.70 on gp41.
To address this discrepancy, the research team evaluated mirror-space cofolding confidence using the Boltz-2 model. For the viral entry target family gp41 (comprising seven structures representing three distinct peptides), interface pLDDT reached a Spearman rho of 0.90 (p = 0.006) and correctly ranked the affinities of all three peptides, outperforming ProteinMPNN NLL (rho = 0.18). Nonetheless, the authors cautioned that because the sample represents only three independent chemotypes, this outcome demonstrates directional consistency rather than a fully validated predictive model.
Future Applications in Antimicrobial Targets
The study also showed that cross-target calibration is currently non-transferable at existing data scales. In addition, the authors proposed prospective design workflows for two critical antimicrobial resistance targets in Pseudomonas aeruginosa: LasR and LecB. All source code, structural data, and benchmarking datasets have been open-sourced on GitHub.