FAR-DPO: Feasibility-Aware and Robust Direct Preference Optimization for Cyclic Peptide Design
Authors: Guofeng Zhang, Rong Han, Xiaoyu Wang, Zhiyun Li, Zongbo Han, Xiaohong Liu, Guangyu Wang
Organizations: Beijing University of Posts and Telecommunications, Beijing, China · Shenzhen University, Shenzhen, China · Tsinghua University, Beijing, China
Cyclic peptides are emerging as promising molecular scaffolds in drug discovery due to their high binding affinity and structural stability. However, extending generative models from linear to cyclic peptide design remains challenging, as cyclization sharply restricts the feasible design space through coupled geometric and biophysical constraints. Moreover, limited training data has led existing approaches to rely largely on zero-shot generation or post hoc filtering, resulting in low yields of feasible designs and limited control over multi-objective trade-offs. To address these limitations, we propose FAR-DPO (Feasibility-Aware and Robust Direct Preference Optimization), an architecture-agnostic framework that steers generative models toward structurally and biophysically feasible cyclic peptide designs, particularly for challenging targets. FAR-DPO integrates feasibility-aware preference construction with difficulty-aware group-robust optimization. Specifically, it constructs within-target preference pairs through feasibility-gated multi-objective dominance and adaptively reweights predefined difficulty groups according to their current preference losses. On the CPSea LNR benchmark, under a fixed generation budget, FAR-DPO increases overall success rate from 46.89% to 57.79% on PepGLAD and from 47.96% to 49.57% on PepFlow. These gains also extend to the hardest target quartile and are accompanied by more favorable best-per-target binding scores. Together, these results demonstrate FAR-DPO's effectiveness in improving feasibility and target-wise robustness.
Figures & tables
Figure 1: Overview of FAR-DPO. (A) Feasibility gates filter generated candidates to form an offline feasible pool. Within each pocket, tolerance-aware dominance is used to construct preference pairs, and frozen pocket-difficulty groups provide their group labels. (B) FAR-DPO is trained on these pairs using group-balanced sampling, reference-relative pair scoring with preferred-candidate anchoring, and loss-adaptive group reweighting. (C) Under identical pockets, sampling budget, and feasibility criteria, the aligned policy increases jointly feasible yield across target-difficulty groups.
Figure 2: Final Success across frozen difficulty groups. The panels use PepGLAD- (top) and PepFlow-defined (bottom) groups; Bars show target-macro results over 14 targets per group, with 95% bootstrap confidence intervals.
Figure 3: Low-budget Final Pass@ K for PepGLAD (top) and PepFlow (bottom). Pass@ K is the target-macro probability of obtaining at least one Final candidate among K samples, evaluated over 56 targets.
Config.
Final ↑
Q4 ↑
Raw@10 Vina ↓
Raw@10 dG ↓
JS ↓
Clash ↓
PepGLAD
46.89
40.00
-5.123
-16.830
0.1596
0.3420
R1: IID/ERM/0
46.14
37.43
-5.604
-16.755
0.1424
0.2805
R2: GB/ERM/0
49.21
43.86
-5.669
-17.554
0.1437
0.2844
R3: GB/ERM/.2
50.32
43.14
-5.578
-17.638
0.1429
0.2832
R5: GB/ GDRO/0
52.77
46.71
-5.726
-16.954
0.1429
0.2549
FAR-DPO
57.79
50.36
−5.579
−17.383
0.1425
0.2592
Table 3: Optimization variants on PepGLAD. Final and Q4 are percentages. Raw@10 Vina and Raw@10 dG denote the target-macro conditional expected-best dG scores, respectively. Configurations use sampler/aggregation/anchor: GB denotes group-balanced sampling and GDRO denotes Group DRO. All preference variants use DPO.
Figure 4: Independent Final-pass designs for 5LM1: (a) PepGLAD, (b) FAR-DPO–PepGLAD; and 3BRH: (c) PepFlow, (d) FAR-DPO–PepFlow. Base/FAR peptides are blue/orange; receptors gray; 4 Å interface residues cyan; cyclization bonds green. Fixed topology-wise best-Vina selection was used.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Backbone
Generated candidates
Final preference pairs
Training preference pairs
Validation preference pairs
PepGLAD
192,000
90,408
85,912
4,496
PepFlow
192,000
92,099
87,494
4,605
Appendix
Table 4: Preference dataset sizes for the two generative backbones.
Setting
PepGLAD
PepFlow
Preference error surrogate
Sequence latent + structure latent
Translation + rotation
Stochastic repeats per preference pair
1
2
β / λanc
15 / 0.2
1 / 0.1
Optimizer / learning rate
AdamW / 1×10−5
Adam / 1×10−5
Adam β1 / β2 / ϵ
0.9 / 0.999 / 1×10−8
0.9 / 0.999 / 1×10−8
Weight decay
0.01
0
Appendix
Table 5: Training settings for PepGLAD and PepFlow.
Method
Final ↑
Q4 ↑
Raw ↓
JS ↓
Clash ↓
Base
46.89
40.00
-5.123
0.1596
0.3420
R1
46.14
37.43
-5.604
0.1424
0.2805
R2
49.21
43.86
-5.669
0.1437
0.2844
R3
50.32
43.14
-5.578
0.1429
0.2832
R4 (IPO)
52.18
46.79
-5.173
0.1595
0.3349
R5
52.77
46.71
-5.726
0.1429
0.2549
Appendix
Table 6: PepGLAD optimization configurations and main results. Final and Q4 are percentages; Raw denotes Raw@10 Vina, JS denotes Interface JS, and Clash denotes Legacy Clash/residue.
Figure 5: (a) Final Success of the PepGLAD optimization variants on Q1–Q4 defined by Base PepGLAD; (b) Base PepFlow and FAR-DPO–PepFlow Final Success on Q1–Q4 defined by Base PepFlow. Points are target-macro means, and error bars are 95% target-bootstrap confidence intervals.
Figure 6: (a) Final Pass@ K under fixed budgets on LNR56; (b–c) fixed- N Matched- N analysis within the PepGLAD and PepFlow backbones. The vertical axes show FAR-DPO-minus-Base differences in expected-best Vina and Rosetta ΔG ; negative values favor FAR-DPO. The analysis uses 56 common targets for N=1,2,3 , 55 for N=5 , and 54 for N=10 ; for N=20 , PepGLAD and PepFlow use 51 and 52 targets, respectively.
Parameter
Value
Final ↑
Vina ↓
Rosetta ΔG↓
Rama allowed ↑
scRMSD Mainchain ↓
Diversity ↑
Novelty ↑
Learning rate η
3×10−6
51.09
-6.75
-23.61
72.9
2.71
0.626
0.122
5×10−6
51.46
-6.75
-23.35
72.5
2.74
0.615
0.123
7×10−6
50.88
-6.80
-23.24
72.2
2.74
0.616
0.123
Group-DRO temperature τ
0.5
53.12
-6.95
-23.26
70.1
2.63
0.610
0.124
2.0
52.89
-6.93
-23.89
69.3
2.60
0.629
0.124
3.0
52.64
-6.92
-23.65
69.7
2.61
0.615
0.122
Appendix
Table 7: PepGLAD training-parameter sensitivity. Final and Rama allowed are percentages, and scRMSD Mainchain is measured in Å. The remaining metrics follow the definitions in the main text. Lower values are better for Vina, Rosetta ΔG , and scRMSD; higher values are better for the other metrics. The final model from the main text is included only as a reference to the main experimental result and is not part of the controlled comparison within each parameter block.
AI Thrust The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China · AI-Peptide Drug Design Joint Laboratory The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China
May 7, 2026·Laura van Weesep, Sunay Chankeshwara, Leonardo De Maria +3Molecular Generation
Molecular AI, Discovery Sciences, Biopharmaceuticals R&D, AstraZeneca, Mölndal, Sweden. · Department of Information Technology, Uppsala University, Sweden · Institute for Complex Molecular Systems (ICMS), Eindhoven University of Technology, The Netherlands. +3