CycWeave: A Rational Token-Free Dual-View Graph Framework for Cyclic Peptidomimetics Modeling

Cyclic peptidomimetics (CPM), combining the developability of small molecules with the targetrecognition capability of larger biomolecules, have emerged as an important molecular modality in drug discovery. Their physicochemical properties and DNA-encoded library (DEL) screening behaviors are jointly governed by local chemical environments, macrocyclic topology, connectivity patterns, and conformational constraints, making them difficult to characterize using traditional molecular representations developed primarily for small molecules. Here, we present CycWeave, a token-free coarse-grained dual-view representation framework tailored to CPM and their complex chemical space of DELs. CycWeave represents each molecule as both an atom-level graph and a fragmentlevel graph at the same time, and replaces discrete tokens with continuous fragment embeddings to jointly model local structural environments and global topological semantics within a unified graph neural architecture. We applied CycWeave to two complementary tasks: CPM membrane permeability prediction and DEL enrichment modeling against transferrin receptor 1 (TfR1). CycWeave achieved robust performance across multiple permeability benchmarks and outperformed general-purpose graph learning models and classical fingerprint-based baselines in a count-driven DEL screening task, supporting its use as a unified representation backbone for CPM property modeling and high-throughput screening modeling.