Machine learning-based stability optimization design and virtual screening of the anti-pulmonary fibrosis peptide...

The novel anti-fibrotic peptide DR3penA, previously developed by our research group, demonstrates substantial therapeutic potential; however, its inherent limited metabolic stability in vivo, attributable to its linear structure, hinders clinical translation. Although cyclization represents a key strategy for enhancing peptide structural rigidity, exhaustive trial-and-error exploration across the vast chemical modification space entails prohibitive computational and experimental costs. In this study, we constructed a machine learning-assisted virtual screening framework for the rational optimization of peptide stability and bioactivity. This framework integrates a stability prediction model with an ensemble bioactivity prediction model based on XGBoost and LightGBM, and employs comprehensive scoring alongside Pareto front analysis to perform multi-objective screening of candidate analogs. Computational evaluation showed that, under leakage-audited cluster splitting, the stability XGBoost model exhibits strong ranking capability (test Spearman ρ = 0.751; NDCG@10 = 0.969), and the XGBoost-LightGBM ensemble achieves high predictive performance for a pan-peptide bioactivity proxy (test R² = 0.822). A virtual library of 95 analogues was generated by predefined chemical rules; cyclization sites were enumerated rather than discovered de novo, and the models were used to rank these schemes. Three analogues (DR3penA_001, _011, and _076) were nominated as a representative panel covering distinct cyclization chemistries, rather than the three highest composite scores. Altogether, this hierarchical, leakage-aware framework enables stability-prioritized, multi-objective nomination of DR3penA analogues and provides a transferable, data-driven strategy for peptide redesign.