CreoPep: A Universal Deep Learning Framework for Target-Specific Peptide Design and Optimization

TL;DR

CreoPep is presented, a deep learning-based conditional generative framework that integrates masked language modeling with a progressive masking scheme to design high-affinity peptide mutants while uncovering novel structural motifs, expanding beyond conventional design paradigms.

Abstract

Target-specific peptides, such as conotoxins, exhibit exceptional binding affinity and selectivity toward ion channels and receptors. However, their therapeutic potential remains underutilized due to the limited diversity of natural variants and the labor-intensive nature of traditional optimization strategies. Here, we present CreoPep, a deep learning-based conditional generative framework that integrates masked language modeling with a progressive masking scheme to design high-affinity peptide mutants while uncovering novel structural motifs. CreoPep employs an integrative augmentation pipeline, combining FoldX-based energy screening with temperature-controlled multinomial sampling, to generate structurally and functionally diverse peptides that retain key pharmacological properties. We validate this approach by designing conotoxin inhibitors targeting the $α$7 nicotinic acetylcholine receptor, achieving submicromolar potency in electrophysiological assays. Structural analysis reveals that CreoPep-generated variants engage in both conserved and novel binding modes, including disulfide-deficient forms, thus expanding beyond conventional design paradigms. Overall, CreoPep offers a robust and generalizable platform that bridges computational peptide design with experimental validation, accelerating the discovery of next-generation peptide therapeutics.

Figures (6)

• Figure 1: Overview of mutation strategies for target-specific peptides.a, $\omega$-Conotoxin features three disulfide bonds. An example is $\omega$-conotoxin MVIIA, which targets Cav2.2 channels (PDBid: 7MIX). b, $\alpha$-Conotoxin contains two disulfide bonds. Examples include $\alpha$-conotoxin ImI and Vc1.1, which primarily target nAChRs (PDBid: 7KOO). c, $\mu$-Conotoxin has three disulfide bonds. An example is $\mu$-conotoxin KIIIA, which blocks voltage-gated sodium channels (PDBid: 6J8E). d, Saturation mutagenesis, where each amino acid residue in the loop1 and loop2 regions of the ImI sequence is replaced by one of the other 19 natural amino acids. Yellow lines indicate disulfide bonds. e, Point mutations are introduced at specific positions identified through alanine scanning, targeting either key or non-key or key residues. f, Target-specific peptide design based on deep learning, directly generating high-potency ImI mutants. g, Schematic diagram of sequence space representation of ImI mutants. The purple dot represents the wild-type peptide, the orange dots represent peptides generated by point mutations, the large purple circle indicates the exploration range of point mutations, the gray dots represent peptides produced by saturation mutagenesis, and the pink dots represent peptides generated by deep learning (DL) methods.
• Figure 2: Overview of CreoPep.a, Schematic representation of the CreoPep framework. Left panel (top to bottom): Training dataset format and feature extraction methods for each component, training phase, and generation phase. Right panel: Mutant screening workflow using FoldX for peptides generated by CreoPep. b, Illustration of three core functional capabilities of CreoPep. c, Data augmentation pipeline used to enhance training diversity and model performance.
• Figure 3: Performance evaluation of the data augmentation pipeline. The x-axis represents different stages of random mutation and data augmentation, while the y-axis shows the $\Delta\Delta G$ values calculated using FoldX. Violin plots display the distribution of $\Delta\Delta G$ values for 1,000 peptide mutants, with the median value represented by the black horizontal line and a dashed horizontal line representing the mean within the box, the lower and upper quartiles delineating the borders of the box, and the vertical black lines indicating the 1.5 interquartile range. A line graph overlays the plots to indicate the average hamming distance between each mutant and the wild-type peptide. For each plot, the left y-axis represents $\Delta\Delta G$ values, the right y-axis corresponds to Hamming distance, and the right panel shows the representative complex structures for each system.
• Figure 4: Classification and generative performance assessment of CreoPep.a, Accuracy of CreoPep in subtype and potency prediction tasks. b, Accuracy of CreoPep in target prediction. c-e, Average pLDDT score, TM-score, and Hamming distance of 1,000 mutants generated by CreoPep. f-h, Comparison of isoelectric point, net charge, GRAVY (hydrophobicity) distributions among the wild-type conotoxin, high potency conotoxins, and 1,000 CreoPep-generated mutants. i-p, Feature distributions in the latent space for the wild-type conotoxin, high potency conotoxins, and 1,000 CreoPep-generated mutants across eight systems.
• Figure 5: Classification and generative performance assessment of CreoPep.a, Bar graph showing the inhibitory effects of 13 candidate conotoxins (at 1 and 10 µM) on ACh-evoked peak current amplitudes mediated by human (h) $\alpha$7 nAChRs. Whole-cell currents were evoked by 100 $\mu$M ACh (mean $\pm$ SD, $n$ = 4-8). The dashed line indicates 50% inhibition of the peak current amplitude. b, Concentration-response relationships for CreoPep-$\alpha$7-1 and CreoPep-$\alpha$7-6 inhibition of ACh-evoked currents at (h) $\alpha$7 nAChRs. Current amplitudes (mean $\pm$ SD; $n$ = 6) were normalized to the response elicited by 100 $\mu$M ACh alone. c, Latent space distribution of 1,000 mutants (gray dots), ImI, high potency $\alpha$7-targeting conotoxins (red dots), low potency $\alpha$7-targeting conotoxins (green dots), and the 13 candidate conotoxins (large yellow dots for high potency; small yellow dots for low potency). The contour plot represents the density distribution of the 1,000 mutants, with darker colors indicating higher density in that region. d, Multiple sequence alignment of ImI with the 13 candidate conotoxins. e, Graphical representation of multiple sequence alignment of 1,000 CreoPep-generated mutants, showing the frequency of amino acid variations at each position relative to the ImI sequence.