acceptodds
Under review as a conference paper at ICLR 2027

PepGuide: Interaction-Guided Latent Diffusion for Target-Conditioned Peptide Design

Abstract

Target-conditioned peptide design poses a significant challenge in drug discovery, primarily because achieving productive peptide–target binding is subject to stringent constraints: the peptide must not only be geometrically plausible, but also establish complementary interactions with the target at the binding interface. However, existing de novo design methods struggle to satisfy these binding-critical constraints simultaneously, resulting in low binding success rates. In this paper, we propose a novel Interaction-Guided Latent Diffusion Framework (PepGuide), which explicitly injects peptide–target interaction information into the generation process to steer peptide synthesis toward conformations that bind effectively to the target. Specifically, we introduce an interaction-aware network (IANet) to capture peptide–target interactions, and integrate the resulting interaction signal into the diffusion model via a time-dependent guidance injection strategy, thereby aligning the guidance with the diffusion trajectory. In addition, we propose a geometry-aware contrastive representation learning to enhance the separability between binder and non-binder samples. Experiments on comprehensive benchmarks show that PepGuide achieves state-of-the-art performance while demonstrating strong effectiveness and robust generalization across multiple binding pockets of a therapeutically relevant target.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Target-conditioned peptide generation, Peptide design, Latent diffusion, All-atom generation.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_wQEfOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PepGuide, a target-conditioned peptide design framework combining a geometry-aware contrastive VAE, an interaction-aware network (IANet), and latent diffusion. IANet is trained to predict Rosetta binding energies from peptide–target complexes. Its interaction features are mapped to a coordinate-space guidance signal, which is injected into the diffusion process according to a time-dependent schedule. At inference, the signal is estimated from an unguided reference peptide selected by VAE perplexity. The authors report results on BioDB and the external LNR benchmark, evaluate multiple binding and structural metrics, and include component and guidance-schedule ablations, a multi-pocket Trop2 study, and an MD analysis on one target.

Critical strengths
  1. The paper addresses a meaningful limitation of target-conditioned peptide generation. PepGuide attempts to translate peptide–target interface information into an explicit signal used during generation, rather than relying only on global target conditioning. The ablation results suggest that both the guidance component and contrastive representation learning contribute to the reported metrics.

  2. The evaluation includes multiple complementary analyses. In addition to sequence and structural metrics, the authors report DockQ and alternative energy scores, compare guidance schedules, and analyze a 50-ns MD trajectory for one target. They also acknowledge that computational evaluation does not replace experimental validation.

Critical weaknesses
  1. The training procedure for the central coordinate-guidance head is not specified clearly enough to reproduce or assess the main mechanism. Section 3.3 defines the guidance signal as Sig=fϕ(Fintp)S_{\mathrm{ig}}=f_\phi(F_{\mathrm{int}}^p), where fϕf_\phi maps IANet features to per-residue coordinate offsets. However, the paper does not state what objective trains fϕf_\phi, whether it is optimized jointly, or whether gradients from the latent-diffusion loss pass through the guided forward process into ϕ\phi. Equation 9 specifies a noise-prediction loss. Appendix F.2, Algorithm 3 freezes the VAE and IANet, initializes the denoiser ϵθ\epsilon_\theta, and updates θ\theta; it does not say whether fϕf_\phi is included in θ\theta, updated separately, or fixed. This ambiguity concerns the mechanism at the center of the contribution.

  2. The independence of the external evaluation and the interpretation of “binding success” need clarification. BioDB is assembled from PepBDB and Q-BioLiP, while LNR is presented as an external test set. The paper describes de-duplication within BioDB and receptor clustering for its internal split, but does not report a cross-source overlap check between LNR and BioDB training data. A PDB-ID overlap check, together with receptor-cluster overlap information, would clarify whether LNR is independent of training. In addition, IANet is trained using Rosetta energy, while the SR column in Table 1 and the Trop2 SR in Table 2 use Rosetta-based criteria. The MD analysis in Appendix E.1 also selects the lowest-Rosetta-energy candidate for each method. DockQ and alternative energy metrics provide useful complementary evidence, but the reported success rates should be interpreted as computational scoring success unless supported by independent evidence.

Questions for authors

Q1: How is fϕf_\phi, the MLP mapping IANet features to coordinate guidance, optimized? Is fϕf_\phi included in the optimizer update for Algorithm 3, and does the gradient from LLDML_{LDM} flow through SigS_{ig} into ϕ\phi? If it is trained separately or fixed, what objective or initialization produces the coordinate-aligned signal?

Q2: Can the authors report the cross-source overlap check for LNR? In particular, were LNR PDB entries or receptor clusters removed from BioDB training? A PDB-ID overlap check and receptor-sequence/cluster comparison would directly address whether the LNR results support external generalization. Please also clarify that SR in Table 1 and the Trop2 SR in Table 2 are defined by the Rosetta criterion.

Reviewer_kk2AOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PepGuide, a new method for generating peptide binders for a specific protein target binding pocket. The main contributions of PepGuide are the introduction of a new Interaction Aware Network, IANet, that captures peptide-target interactions, and a new guidance method that uses interaction information from IANet to help steer generated designs to have improved binding interactions with the target protein. The authors validate this approach by doing retrospective evaluation on protein target and peptide complexes from their own curated dataset, BioDB, and on an existing previously reported curation of protein-peptide complexes called LNR, showing results that are better than other previously reported methods.

Critical strengths

This paper is very thorough, and provides good mathematical definitions and descriptions of its key innovations. The overall goal of developing new methods for generating peptide binders for specific target binding pockets, and integrating in additional interaction information, is also strongly motivated.

Critical weaknesses

One primary weakness of this paper is in its evaluation metrics.

First, no statistical comparisons, error bars, or uncertainties are reported for any of the evaluations or methods. Given that many of the models show similar performance, it is difficult to see if PepGuide's performance improvements are statistically significant.

Second, many of the metrics reward similarity to the single native peptide in the original test protein-peptide complex. Metrics like AAR and receptor aligned C_alpha RMSD, and DockQ all reward similarity to the reference native peptide, even though one of the goals of a generative model is the ability to sample many potential peptide binder designs. Some of these goals directly clash. High AAR sequence recovery and sequence diversity cannot both be maximized by the same model, just like recovering native C_alpha RMSD and structural diversity. This makes the value of getting the highest or lowest scores in each of these categories unclear.

Additionally, despite the paper emphasizing sequence-and-structure co-design, the metrics for physical validity only focus on C_alpha backbones, which ignores many potential clashes in the side-chains that this approach explicitly designs. This metric is also somewhat unclear since in practice, many methods (see AlphaFold3 or cofolding models) apply relaxation steps to fix local geometry, which can often be repaired.

However, the most important questions are with the binding quality metrics. Two key metrics, success rate and the BNR_Ros rate are both explicitly calculated based on Rosetta binding energies, which are also directly used to supervise IANet in the training of PepGuide. This makes using this metric as an independent assessment of generated peptides confounded, since the other methods do not receive this supervision. Given that Rosetta binding energy is a very imperfect oracle for real world binding affinity, this raises the question of whether the scores show real binding improvements, or just a trained bias towards Rosetta favored designs that inflates their scores. While the included MM-GBSA metric and EvoEF2 metrics partially address this, the MM-GBSA methodology is unclear and MM-GBSA varies wildly in accuracy, with high sensitivity to initial conditions, and EvoEF2 is one method among many other widely used metrics for assessing peptide designs. In addition to EvoEF2, it would be helpful if the authors could justify the performance improvements via other methods that are common in the field, such as reporting HADDOCK protein-peptide docking results or AlphaFold3 folded complex quality, as additional ways to assess each model's performance with a variety of different complementary computational oracles. Otherwise, the metrics are very difficult to assess for quality and likelihood of binding.

The second primary weakness of this paper is in its comparison methods. First, this paper does not cite or compare to RADiAnce, which is another peptide design method that explicitly integrates a contrastive pocket conditioned objective. This paper also does not cite or compare to DiffPepBuilder or PPFlow, which are additional methods that target this same problem, and have been cited previously in the field.

More broadly, many other methods have demonstrated experimental success for the design of peptide binders. These include BindCraft or RFdiffusion3. While some of these do not do full sequence-and-structure co-design, as this paper points out, in practice other tools such as ProteinMPNN allow them to be easily applied to the full design of peptide binders.

Some of these methods may not be comparable for various reasons to PepGuide, and the authors definitely don't need to benchmark against all possible methods. However, a strong justification for scoping why certain methods are or are not compared against given that many have been adopted for peptide design by others or demonstrated experimental success is important.

Questions for authors

What was the protocol for running molecular dynamics, and were replicates run to demonstrate the stability results hold?

What was the specific procedure for calculating each of the metrics displayed in Table 1, especially metrics that require alignment or follow-up procedures such as MM-GBSA or DSSP based SSR?

Did all baselines get trained on the same training data, including the ProtFrag dataset? Was the ProtFrag data also clustered and split to avoid leakage with BioDB and LNR test complexes?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.