EDWARD: E(3)-Equivariant Dual-Way Attentive Reduction for Peptide-to-Small-Molecule Design

Peptides, valued for their high affinity and selectivity yet hindered by suboptimal pharmacokinetics, are important entry points for lead discovery; the canonical strategy preserves their key binding interactions and prunes them into developable small molecules. However, pharmacophore-based simplification methods heavily depend on high-quality databases and expert knowledge, and pocket-conditioned generation alone struggles to align critical pharmacophores and orientations, limiting the controllability and efficiency. We therefore introduce EDWARD, a geometry-pharmacophore joint framework that learns "pharmacophore cloud clusters" and radial-directional constraints from peptide-protein complexes, and performs geometry-first assembly of modular fragments and multianchor scaffolds, thereby avoiding end-to-end scaffold learning. This strategy stably maintains alignment and pocket compatibility across conformations, lengths, and topologies; achieves a better balance among affinity retention, drug-likeness, and diversity; and requires little annotation with controllable computational cost, providing a supplementary route from peptide cues to translatable small molecules.