FAR-DPO Boosts Feasibility in Cyclic Peptide Design

This research introduces FAR-DPO, a novel framework designed to overcome significant challenges in extending generative models from linear to cyclic peptide design for drug discovery. Cyclic peptides face stringent geometric and biophysical constraints, making their design space highly restricted and leading to low yields of feasible designs with existing methods. FAR-DPO addresses these limitations by steering generative models towards structurally and biophysically feasible cyclic peptide designs, particularly for difficult targets. It achieves this through two main mechanisms: feasibility-aware preference construction, which builds within-target preference pairs using feasibility-gated multi-objective dominance, and difficulty-aware group-robust optimization, which adaptively reweights predefined difficulty groups based on their current preference losses. Evaluations on the CPSea LNR benchmark demonstrate FAR-DPO's effectiveness. Under a fixed generation budget, it increased the overall success rate for PepGLAD from 46.89% to 57.79% and for PepFlow from 47.96% to 49.57%. These improvements were also observed in the hardest target quartile, alongside more favorable binding scores, showcasing its ability to enhance both feasibility and target-wise robustness.