Year: 2026
Type: preprint
Abstract: Macrocycles and cyclic peptides are increasingly important in beyond-rule-of-five (bRo5) drug discovery, yet their absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties remain difficult to predict owing to sparse data, complex topology, and environment-dependent conformations. Here, we present AIO (All-in-One), an interpretable multimodal ADMET framework combining topology-preserving molecular segmentation with solvent-aware contrastive pretraining. Across repeated cross-validation, AIO performed strongly across permeability, solubility, unbound fraction, clearance, half-life, bioavailability, and toxicity endpoints, achieving Pearson R = 0.87 for PAMPA permeability and AUROC = 0.91 for bioavailability. Case studies of recently approved macrocyclic and cyclic-peptide drugs showed agreement between predictions and reported pharmacokinetic profiles, while segment-level attribution highlighted molecular regions associated with predicted ADMET bottlenecks. Despite pretraining exclusively on macrocycles, AIO also transferred competitively to the small-molecule PharmaBench dataset. These results establish a promising foundation for interpretable ADMET prediction across diverse molecular modalities in drug discovery.
Authors: Sicheng Wen, Shu Wang, Yue Qian
Institution: Viva Biotech (China)
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7236579034
