NS-Pep: Unified NSAA-Aware Peptide Design

• NS-Pep is a unified framework for de novo peptide design that integrates non-standard amino acids (NSAAs) with pocket-conditioned folding.
• It employs Residue Frequency-Guided Modification, Progressive Side-chain Perception, and Interaction-Aware Weighting to address data imbalance and improve side-chain modeling.
• Empirical results demonstrate significant improvements in sequence recovery, binding affinity, and structural accuracy compared to standard peptide generators.

NS-Pep is a unified framework for de novo peptide sequence–structure co-design and pocket-conditioned peptide folding with non-standard amino acids (NSAAs). It is motivated by the observation that peptide drugs incorporating NSAAs can improve binding affinity and pharmacological properties, while existing peptide design methods are largely restricted to standard amino acids (SAAs). The framework addresses three coupled difficulties: the severe long-tailed distribution of residue types, insufficient side-chain modeling for NSAAs, and the lack of a unified NSAA-aware folding model with pocket conditioning. Its central components are Residue Frequency-Guided Modification (RFGM), Progressive Side-chain Perception (PSP), and Interaction-Aware Weighting (IAW) (Guo et al., 1 Oct 2025).

1. Problem setting and scope

NS-Pep is formulated for peptide design in which the generated binder is conditioned on a protein pocket and may contain NSAAs such as SEP, PTR, and TYS. The paper identifies a fundamental data-imbalance problem: NSAAs are extremely underrepresented, and even the most frequent one, SEP, accounts for less than $0.4\%$ of residues, while the 18 most frequent NSAAs together account for only about $2\%$ of residues. This creates a regime in which naive training is dominated by SAAs, and standard cross-entropy optimization can suppress rare residue classes too strongly (Guo et al., 1 Oct 2025).

The method is also positioned against structural limitations of earlier peptide generators. Many prior systems model side chains only through torsion angles, which is too coarse when NSAAs differ from standard residues by subtle atomic modifications. In parallel, the paper distinguishes between tools that support NSAAs without pocket specification and tools that support pocket-conditioned folding but not NSAAs, and frames NS-Pep as filling that gap.

The model’s design target is a peptide represented residue-wise by

$\{a^j, x^j, R^j\}_{j=1}^n,$

where $a^j$ is residue type, $x^j \in \mathbb{R}^3$ is backbone translation, and $R^j \in \mathbb{R}^{3\times 3}$ is backbone rotation, together with side-chain variables

$\chi^j \in [0,2\pi)^7.$

The conditioning input is a protein pocket

$$\mathcal{I}^{\text{poc}}=\{a^j, x^j, R^j, \chi^j\}_{j=1}^m.$$

2. Generative formulation

NS-Pep is built on conditional flow matching over manifolds for residue identities, backbone translations, and backbone rotations, and then extends this backbone-and-sequence generation stage to explicit side-chain prediction. The main generation objective is written as

$$\mathcal{L}_{\text{CFM}}(\theta)= \mathbb{E}_{t,p_1(h_1),p_0(h_0)} \left\|v_{t,\theta}(h_t,t)-u_t(h_0,h_1,t)\right\|^2,$$

where $v_{t,\theta}$ is the learned time-dependent vector field and $2\%$0 is the target conditional vector field. At inference time, sampling is performed by integrating

$2\%$1

This defines the core sequence–backbone co-design process (Guo et al., 1 Oct 2025).

A distinctive aspect of the framework is that side-chain prediction is not coupled from the beginning to the noisiest stage of generation. Instead, the model first refines residue identity and backbone geometry, and only later activates the side-chain branch. Full all-atom reconstruction uses an AlphaFold2-style atom placement map

$2\%$2

This separation reflects the paper’s claim that side-chain prediction should be conditioned on a sufficiently stable residue and backbone state.

This suggests a two-level generative factorization: first, peptide identity and backbone are generated under pocket conditioning; second, side-chain geometry is inferred conditional on the generated backbone and residue types. A plausible implication is that the framework is intended to reduce ambiguity that would arise if torsions were predicted while residue identity remained uncertain.

3. Residue Frequency-Guided Modification

RFGM is the component that targets long-tail learning over residue types. Its purpose is to mitigate over-penalization of rare amino acids by modifying logits in a residue-frequency-aware way. For class $2\%$3, the modified logit is defined as

$2\%$4

where $2\%$5 is the count of class $2\%$6 and $2\%$7 (Guo et al., 1 Oct 2025).

The paper motivates this through the standard cross-entropy objective

$2\%$8

with gradient

$2\%$9

For a negative class, the gradient reduces to $\{a^j, x^j, R^j\}_{j=1}^n,$0, so repeated exposure to head-class samples can drive rare-class logits downward. RFGM alters the effective gradients so that tail classes receive smaller negative penalties and head classes receive stronger ones.

The paper also provides a theoretical justification. It states that if a nonnegative perturbation is added to a logit, the gradient magnitude with respect to that class is monotonic in the direction of the perturbation. In a two-class setting, the tail class has a probability greater than $\{a^j, x^j, R^j\}_{j=1}^n,$1 of experiencing reduced gradient magnitude after modification. In the multi-class setting, the analysis introduces

$\{a^j, x^j, R^j\}_{j=1}^n,$2

and argues that more frequent classes have larger expected perturbation terms, increasing their negative gradient penalties relative to tail classes.

Empirically, the long-tail effect is large. On the NSAA test set, PepFlow* has $\{a^j, x^j, R^j\}_{j=1}^n,$3, PepFlow* + RFGM reaches $\{a^j, x^j, R^j\}_{j=1}^n,$4, and the full NS-Pep system reaches $\{a^j, x^j, R^j\}_{j=1}^n,$5. The paper also reports that removing the noise component from RFGM causes a clear drop in AAR, indicating that the perturbation term is not merely incidental regularization.

4. Progressive Side-chain Perception and interaction-aware optimization

PSP is introduced to address the claim that torsion-only side-chain modeling is insufficient for NSAAs. The method decouples generation into a sequence-plus-backbone stage and a side-chain stage conditioned on the generated residue type and refined backbone. Side-chain prediction is activated only when

$\{a^j, x^j, R^j\}_{j=1}^n,$6

so the model predicts torsions and atom-level corrections only in the later denoising regime, when the backbone is more stable (Guo et al., 1 Oct 2025).

The PSP loss is given residue-wise by

$\{a^j, x^j, R^j\}_{j=1}^n,$7

where

$\{a^j, x^j, R^j\}_{j=1}^n,$8

The first term supervises torsion prediction; the second supervises local atom-level geometric correction. The paper’s interpretation is that torsions provide only coarse geometry and that atom-level offsets are needed to capture distinctions between standard and non-standard side chains.

IAW weights losses according to proximity to the receptor pocket. For peptide residue $\{a^j, x^j, R^j\}_{j=1}^n,$9, the weight is

$a^j$0

where $a^j$1 is the distance between the closest atom pair of peptide residue $a^j$2 and receptor residue $a^j$3, and $a^j$4 is set to $a^j$5 Å. These weights are incorporated into the total objective,

$a^j$6

The stated rationale is that residues closer to the pocket are more important for binding and conformational stability.

Taken together, PSP and IAW make the geometric objective both more detailed and more localized. This suggests that NS-Pep is not only a residue-type generator but also a binding-interface-aware structure model.

5. Data, baselines, and evaluation protocol

The benchmark combines Q-BioLip and PepBDB. It contains 10,348 complexes with peptide length $a^j$7–$a^j$8, split into 8,828 training complexes, 1,065 validation complexes, and 219 test complexes. The paper distinguishes a General test set from an NSAA test set that contains only complexes with at least one NSAA. In the main paper, the NSAA setting focuses on SEP, TYS, and PTR; in the appendix, the alphabet is extended to 38 residue types, comprising 20 SAAs and 18 NSAAs with at least 30 occurrences each (Guo et al., 1 Oct 2025).

Training uses distributed data parallel on 2 NVIDIA A100 GPUs for 320k steps with learning rate $a^j$9. Validation uses 100 sampling timesteps, generation uses 200 sampling timesteps, and 16 peptide samples are generated per test case. The IAW threshold is $x^j \in \mathbb{R}^3$0 Å.

The reported baselines for peptide generation are RFdiffusion, ProteinGenerator, PPFLOW, DiffPepBuilder, PepGLAD, PepFlow, and PepFlow*. For long-tailed learning comparisons, the paper evaluates Weighted Sampling, BLV, and Seesaw loss. For folding, it compares against ESMFold, AlphaFold3, and PepFlow.

Evaluation covers three dimensions. Recovery metrics are AAR, AAR(S), RMSD, SSR, and BSR. Energy is summarized by AFF, defined as the percentage of generated peptides with better Rosetta binding affinity than the native binder. Design consistency and diversity are measured by scRMSD and Diversity. For folding, the paper additionally reports Success, defined as RMSD $x^j \in \mathbb{R}^3$1 Å, and AFF(Success).

6. Empirical results, interpretation, and limitations

On the General test set for de novo peptide design, NS-Pep reports $x^j \in \mathbb{R}^3$2, $x^j \in \mathbb{R}^3$3, $x^j \in \mathbb{R}^3$4 Å, and $x^j \in \mathbb{R}^3$5 Å. Relative to the second-best methods, the paper states improvements of $x^j \in \mathbb{R}^3$6 in sequence recovery and $x^j \in \mathbb{R}^3$7 in binding affinity. It also notes that PepFlow*, which is simply adapted to NSAA generation, performs worse than the original PepFlow on AAR, and uses this as evidence that naive NSAA extension is insufficient (Guo et al., 1 Oct 2025).

On the pocket-specific folding task, the reported results are as follows:

Method	Metric	Reported value
AlphaFold3	Success	10.49
PepFlow	Success	25.52
NS-Pep	Success	28.25
AlphaFold3	RMSD	25.30
PepFlow	RMSD	4.96
NS-Pep	RMSD	4.90

The paper summarizes these results by stating that NS-Pep outperforms AlphaFold3 by $x^j \in \mathbb{R}^3$8 in peptide folding success rate and by $x^j \in \mathbb{R}^3$9 Å in RMSD. It further reports $R^j \in \mathbb{R}^{3\times 3}$0 and $R^j \in \mathbb{R}^{3\times 3}$1 for NS-Pep in folding.

The ablation study on the NSAA test set attributes distinct roles to the three modules. RFGM primarily improves NSAA recovery; PSP improves recovery, structure, and affinity; IAW improves affinity by emphasizing pocket-proximal residues. In the full system, the paper reports $R^j \in \mathbb{R}^{3\times 3}$2, $R^j \in \mathbb{R}^{3\times 3}$3, and $R^j \in \mathbb{R}^{3\times 3}$4. In the expanded 18-NSAA experiment, PepFlow* reaches $R^j \in \mathbb{R}^{3\times 3}$5, whereas NS-Pep reaches $R^j \in \mathbb{R}^{3\times 3}$6, with reported per-residue gains including SEP $R^j \in \mathbb{R}^{3\times 3}$7, PTR $R^j \in \mathbb{R}^{3\times 3}$8, ALY $R^j \in \mathbb{R}^{3\times 3}$9, DLY $\chi^j \in [0,2\pi)^7.$0, SAR $\chi^j \in [0,2\pi)^7.$1, and MVA $\chi^j \in [0,2\pi)^7.$2.

The paper explicitly notes several limits. Ultra-rare NSAAs with fewer than 10 samples remain difficult to generalize to, and few-shot learning is suggested as a possible remedy. Broader drug-like properties such as bioavailability and membrane permeability are not evaluated. The authors also state that further progress may require few-shot learning, reinforcement learning, and additional experimental validation. A separate concern noted in the paper is misuse for harmful peptide design.

In this formulation, NS-Pep is best understood not as a simple residue-vocabulary extension of standard peptide generators, but as an NSAA-aware co-design and folding model whose improvements derive from long-tail calibration, side-chain-specific geometric refinement, and interface-aware weighting. That interpretation is consistent with the empirical result that the largest gains appear precisely where naive extension fails: rare residue recovery, side-chain-sensitive structure generation, and pocket-specific folding (Guo et al., 1 Oct 2025).