NS-Pep: De novo Peptide Design with Non-Standard Amino Acids

NS-Pep is a unified flow-matching framework for co-designing peptide sequences and structures that incorporate non-standard amino acids (NSAAs). It introduces Residue Frequency-Guided Modification (RFGM), Progressive Side-chain Perception (PSP), and Interaction-Aware Weighting (IAW) to tackle the long-tailed NSAA distribution and to capture fine-grained side-chain geometry.

Peptide Generation on the General Test Set

Table 1 of the paper: comparison for peptide generation on the General test set.

Method	AAR % ↑	AAR(S) % ↑	RMSD Å ↓	SSR % ↑	BSR % ↑	AFF % ↑	scRMSD Å ↓	Diversity ↑
RFdiffusion	8.66	8.69	17.57	71.75	6.52	12.06	19.64	0.41
ProteinGenerator	7.08	7.12	17.26	75.30	4.28	8.91	23.02	0.39
PPFLOW	8.63	8.87	9.45	0.11	75.14	6.53	15.77	0.66
DiffPepBuilder	6.71	6.77	8.36	75.49	85.86	12.11	14.65	0.51
PepGLAD	6.82	6.90	7.78	78.72	81.48	10.36	13.43	0.72
PepFlow	16.30	16.76	4.98	85.80	81.50	11.60	12.32	0.50
PepFlow*	15.88	15.88	4.94	83.40	84.36	11.39	11.67	0.51
NS-Pep (ours)	22.53	22.35	4.92	85.00	81.62	17.23	11.50	0.49

Bold = best; italic = second best.

Model Weights & Evaluation

The checkpoint and evaluation files below reproduce the NS-Pep row of the table above (AAR = 22.53%).

Item	Path
Checkpoint (320k iterations; model weights + config only)	checkpoints/NS-Pep_320k.pt (Git LFS)
Model class	models_con/flow_model.py → NS_Pep
Evaluation code	eval_recovery.py
Evaluation script	scripts/eval_NS3.sh
Config (used to train this checkpoint)	configs/NS-Pep_320k.yaml
NSAA constants (SEP / TYS / PTR)	pepflow/modules/protein/constants.py
Dataset split (TRAIN / VAL / TEST = 8228 / 1065 / 219)	data/data_split_NS3_v3.json
Reference results (16 samples per test case)	results/test_320k_num16/eval_recovery.csv
Environment (bio)	environment/bio_environment.yml, environment/bio_conda_list.txt, environment/bio_pip_freeze.txt

Data

The dataset can be downloaded from Google Drive: https://drive.google.com/file/d/1j459qA_2MFAvEHaZ6CceT5B8a3z8HKw0/view?usp=sharing

After unzipping, each test case is a folder named by its ID (e.g. 1a46_I/). It contains the receptor/pocket/peptide PDB files and the preprocessed <ID>_re_NS3.pt. If that .pt file is missing, eval_recovery.py rebuilds it from peptide_re.pdb / pocket_re.pdb.

Run

git lfs pull                          # fetch the checkpoint
conda env create -f environment/bio_environment.yml
conda activate bio
bash scripts/eval_NS3.sh /path/to/Pepflow/ ./outputs/test_320k_num16

The script writes eval_recovery.csv to the output directory, with one row per test case and an All summary row: rec_seq = AAR, rec_pos = RMSD, rec_ss = SSR, rec_bs = BSR. The Clash column is disabled and always reads 100. Sampling is stochastic, so a fresh run gives a slightly different value. Our re-run on a V100 gave an AAR of 22.78%, compared with the 22.53% reference.

Notes:

• eval_recovery.py and the config files still contain our cluster paths as defaults. scripts/eval_NS3.sh overrides every path that inference needs.
• The model class is selected by the uncommented model = flow_model.NS_Pep(...) line in eval_recovery.py.

Citation

@article{guo2025ns,
  title={NS-Pep: De novo Peptide Design with Non-Standard Amino Acids},
  author={Guo, Tao and Yin, Junbo and Wang, Yu and Gao, Xin},
  journal={arXiv preprint arXiv:2510.03326},
  year={2025}
}