PepEDiff: Zero-Shot Peptide Binder Design via Protein Embedding Diffusion

TL;DR

PepEDiff is presented, a novel peptide binder generator that designs binding sequences given a target receptor protein sequence and its pocket residues, and leverages the global protein embedding manifold as a semantic prior, allowing the model to propose novel peptide sequences in previously unseen regions of the protein space.

Abstract

We present PepEDiff, a novel peptide binder generator that designs binding sequences given a target receptor protein sequence and its pocket residues. Peptide binder generation is critical in therapeutic and biochemical applications, yet many existing methods rely heavily on intermediate structure prediction, adding complexity and limiting sequence diversity. Our approach departs from this paradigm by generating binder sequences directly in a continuous latent space derived from a pretrained protein embedding model, without relying on predicted structures, thereby improving structural and sequence diversity. To encourage the model to capture binding-relevant features rather than memorizing known sequences, we perform latent-space exploration and diffusion-based sampling, enabling the generation of peptides beyond the limited distribution of known binders. This zero-shot generative strategy leverages the global protein embedding manifold as a semantic prior, allowing the model to propose novel peptide sequences in previously unseen regions of the protein space. We evaluate PepEDiff on TIGIT, a challenging target with a large, flat protein-protein interaction interface that lacks a druggable pocket. Despite its simplicity, our method outperforms state-of-the-art approaches across benchmark tests and in the TIGIT case study, demonstrating its potential as a general, structure-free framework for zero-shot peptide binder design. The code for this research is available at GitHub: https://github.com/LabJunBMI/PepEDiff-An-Peptide-binder-Embedding-Diffusion-Model

Figures (6)

• Figure 1: Overview of the structure-free peptide design framework.a) We adopt the diffusion framework (top row), which iteratively refines a random signal ($x_T$) into a final peptide embedding ($x_0$). The reverse denoising process, $p_\theta(x_{t-1}|x_t, z, m)$, is conditioned on the target receptor at each step. The denoising network (bottom row) employs a cross-attention mechanism to focus on the receptor's binding pocket, represented by its embedding $z$ and a pocket mask $m$. The final refined embedding $x_0$ is then decoded to generate the binder sequence $\hat{S}$. The generated sequence is subsequently used to predict the receptor–peptide complex structure. b) An illustration of out-of-distribution sampling. By exploring the latent space of all know protein sequence, we aim to generate peptides out of the relative small know peptide binder distribution.
• Figure 2: Performance comparison with baseline methods. Top row shows the sequence and structure diversities. Our model achieves superior diversity in both sequence and structure comparing to the two baseline model. Bottom row shows the distribution of generated peptides embedding (dimension reduced using UMAPmcinnes2018umap) comparing to the testing set (shown as gray area). Major distribution differences are marked with color box.
• Figure 3: Generated TIGIT binders. The top row shows some generated peptides for all baseline method. For the RF&MPNN pipeline, all ten generated structures are visualized. For PepEDiff and DiffPepBuilder, ten peptides (out of 100) are shown for visual clarity. The bottom row shows a combined Ramachandran plot for all generated peptide structures, and the dissimilarity ($1-\text{similarity}$) between generated peptide and training set.
• Figure 4: Top-Performing Binders for the TIGIT Receptor. The three leftmost figures show the docking poses. The number of repeats showing interactions between the peptide and pocket is marked beside the method. The rightmost figure shows the energy distribution of generated peptides. The exact sequence of top-performing binder for our model, RF&MPNN, and DiffPepBuilder are $LRISSDVHQDAASVH$, $SRAEQNAALLARVAG$, and $ILDDILRAAALAAGF$, respectively. All 100 generated binders are provided in our GitHub repository.
• Figure 5: Docking Simulation Analysis. a) van der Waals (vdW) interaction energy between the peptide and TIGIT during the simulation. Our peptide exhibits the earliest interaction with TIGIT and also shows the strongest (lowest) interaction energy. b) Root-mean-square deviation (RMSD) of the peptide over the course of the simulation. All three peptides reach a stable conformation after approximately 800 ns. c) Root-mean-square fluctuation (RMSF) of the peptide throughout the simulation.