DEL-Flow: AI-Driven Structural Generation and Activity Validation of MDM2-Targeting Macrocyclic Peptides

Macrocyclic peptides represent a promising modality for modulating intracellular protein–protein interactions; however, macrocyclic peptide datasets with reliable experimental activity measurements that can inform rational design remain limited. Here, we developed DEL-Flow, a pocket-conditioned generative framework for the design of MDM2-p53 inhibitors. Using an iterative DEL-selection strategy, a first-generation symmetric dimeric macrocyclic peptide DEL identified a half-scaffold associated with highly enriched library members, which was retained to construct a second-generation asymmetric focused DEL. Experimentally derived selection signals from this focused DEL were combined with computationally generated conformations and macrocyclic topologies for model training. DEL-Flow integrates these priors with MDM2 pocket and PackDock-derived pose information to generate side-chain variants while retaining the macrocyclic core. An SE(3)-equivariant flow-matching model generates all-atom coordinates, and a MASK denoising module predicts editable-region atom types, chirality, and bond connectivity. Starting from templates 126 and 319, DEL-Flow generated 20 candidates from each template, yielding 40 candidates in total. Following prioritization, 126-18 and 319-18 were synthesized off DNA and inhibited the MDM2-p53 interaction in fluorescence-polarization assays, with IC50 values of 1.95 and 6.50 μM and Ki values of 163 and 548 nM, respectively. These results support the use of DEL selection signals and pocket geometry to guide the generation of macrocyclic peptides absent from the original library that inhibit the MDM2–p53 interaction.