CP-GFM-SGX: Topology-Aware Cyclic Peptide Generation and Design with Receptor-Conditioned Structural Evaluation and Experimental KRAS Affinity Optimization

Abstract

Cyclic-peptide discovery requires molecular topology, atomic structure, receptor-associated evidence and computational decisions to remain consistent across a multistage workflow. We present CP-GFM-SGX, an all-atom generative AI model for cyclic-peptide design that integrates a topology-aware generative core with a Unified Evidence State Machine (UESM) to track candidate identity, chemical validity, provenance, typed failures and bounded computation. We evaluated three distinct aspects of the system. In Track A, a frozen five-complex known-ligand benchmark with 100 scheduled poses per complex, the native-free FULL route selected a ligand-backbone pose with root-mean-square deviation (RMSD) below 2 Å in four of five systems; retrospective best-of-100 evaluation found a sub-2-Å pose in all five (median selected RMSD, 1.72 Å). In Track B, matched true-pocket versus shuffled-pocket comparisons favored the specified pocket in all seven main targets (median of target medians, +172.6 Rosetta energy units; weakest margin, +3.9 for MCL-1–9CDT). An exploratory post-protocol assessor admitted 42/42, 42/42 and 40/42 selected top-K slots in the full, condition-masked and post hoc head-to-tail arms, respectively. In Track C, FULL UESM produced the specified protective outcomes for six deterministic synthetic fault fixtures and one additional missing-evidence case; targeted ablations isolated the roles of provenance and failure propagation. In a separate KRAS G12D–GDP experimental-feedback series, synthesis and SPR/BLI testing of cyclic-peptide analogues guided position-2 basic-side-chain selection and dual-basic-group placement, advancing the reported SPR affinity from approximately 60 µM for CRO-748 to approximately 6 µM for 748-R1 and approximately 1 µM for 748-RK. Comparisons using the 31-dimension Structured Capability and Evidence Grid (SCEG-31D) demonstrate that CP-GFM-SGX exhibits certain superiority over other design platforms in multiple aspects. Together, we evaluate a topology-aware generative artificial intelligence workflow for receptor-focused cyclic-peptide design and show, in a KRAS feedback series, how binding measurements guide iterative chemical optimization.