A narrow thermodynamic design window governs selective membrane permeabilization and antiviral activity of amphipathic peptides

Designing molecules that selectively target therapeutically relevant membranes, such as viral envelopes, while sparing host cells is challenging: these membranes closely resemble host bilayers, so selectivity must exploit subtle lipid composition and curvature differences and demands precise tuning of affinity and hydrophobicity, yet curated sequence-specificity data are scarce. Here we show that selective membrane permeabilization and membrane-selective activity of amphipathic peptides are governed by a narrow thermodynamic design window defined by membrane curvature affinity and molecular hydrophobicity. Using a physics-driven generative workflow combining evolutionary molecular dynamics and a transformer predictor (PMIpred), we systematically explored and thermodynamically mapped peptide sequence space de novo without reliance on natural templates or experimental training data. Across four design generations we synthesized and experimentally characterized 43 peptides. Mapping functional activity onto a low-dimensional free-energy landscape reveals a confined thermodynamic "sweet spot" separating weak membrane binding from excessive hydrophobic association and cytotoxicity. Peptides operating within this regime efficiently permeabilize model membranes while maintaining low cellular toxicity. Antiviral activity against Zika virus and HIV-1 emerges in the same region but depends sensitively on membrane lipid composition. Quantitative thermodynamic design rules emerge for membrane-active peptides, illustrating how low-dimensional free-energy landscapes can guide the engineering of selective interactions at soft-matter interfaces.