Exploring a Structural Data Mining Approach to Design Linkers for Head-to-Tail Peptide Cyclization

Yasaman Karami, Samuel Murail, Julien Giribaldi, Benjamin Lefranc, Florian Defontaine, Olivier Lesouhaitier, Jérôme Leprince, Sjoerd Jacob de Vries, Pierre TufféryJournal of Chemical Information and Modeling 2023 · OpenReview

TL;DR

This paper introduces PEP-Cyclizer, a novel computational framework for designing head-to-tail cyclization linkers in peptides. By leveraging structural data mining, PEP-Cyclizer can either predict the 3D conformation of a cyclized peptide given a linker sequence or propose effective linker sequences for a given linear peptide, significantly advancing rational cyclic peptide drug design.

What Problem Does This Paper Address?

Peptides are promising therapeutic candidates due to their high affinity, target selectivity, and low toxicity. However, their development is often hindered by poor bioavailability and susceptibility to enzymatic degradation. Head-to-tail cyclization, a strategy that forms a covalent bond between the C-terminus and N-terminus of a peptide, can enhance stability and improve pharmacokinetic properties. While this approach is powerful, designing effective linkers that connect these termini without disrupting the peptide's native conformation or activity has been a significant challenge. Existing computational methods often focus on de novo design of entire cyclic peptides or require extensive experimental validation, leaving a gap for a method that can rationally design linkers for existing linear peptides.

Methods

PEP-Cyclizer employs a multi-step workflow:

  1. Input Preparation: Takes a linear peptide structure and either linker sequence constraints or a linker sequence for conformation modeling.
  2. Dataset Search: Mines a structural database (PDB) to identify potential linker candidates that fit the geometry of the peptide's N- and C-terminal "flanks" (defined as four residues before and after the termini).
  3. Filtering: Applies several filters to narrow down candidates:
  4. Sequence Similarity: Ranks candidates based on similarity to a query linker sequence using BLOSUM62 scores.
  5. Geometrical Clustering: Groups similar linker candidates using hierarchical clustering based on RMSD.
  6. Local Conformation: Measures the similarity of local conformation profiles between query and candidate linkers using Jensen-Shannon divergence.
  7. Steric Clash Detection: Discards models with steric clashes.
  8. Model Building: For conformation modeling, it generates and refines 3D models of the cyclized peptide. For sequence prediction, it uses a forward-backtrack algorithm to infer likely linker sequences.
  9. Ranking/Logo Generation: Ranks predicted linker sequences or generates sequence logos for visualization.

A key novelty is treating head-to-tail cyclization as a loop-modeling problem, allowing the use of existing structural databases and algorithms adapted from loop prediction. The method also includes a formula to estimate the minimum linker size based on the head-to-tail distance.

Datasets

  • Protein Data Bank (PDB): Used as the primary structural database for identifying linker candidates and for evaluating model performance.
  • CyBase: A database of natural cyclic peptides, used in conjunction with PDB data to establish relationships between residue separation and inter-residue distance for predicting minimum linker size.
  • Conotoxin Benchmark: A set of seven cyclized/non-cyclized peptide pairs from the PDB used to validate the linker modeling and sequence prediction capabilities.
  • Nrf2 and Urotensin II (UII) peptides: Used as case studies for applying PEP-Cyclizer to design functional cyclic peptides.

Key Insights and Ideas

  • Head-to-tail cyclization as loop modeling: This conceptual shift allows leveraging existing structural data and algorithms effectively.
  • Robustness to conformational inaccuracies: PEP-Cyclizer performs well even when the starting linear peptide structure has some inaccuracies, a crucial advantage for real-world applications.
  • Predictive power for linker sequences: The method can rank experimental linker sequences significantly better than random ones, suggesting it captures relevant sequence-structure relationships.
  • Linker size estimation: The development of a formula to estimate the minimum linker size based on head-to-tail distance provides a practical starting point for design.
  • Experimental validation confirms utility: The successful design and synthesis of cyclized Nrf2 and UII peptides with improved or retained activity demonstrate the practical applicability of PEP-Cyclizer.

Key Results

  • Conformation Modeling: PEP-Cyclizer achieved an average global RMSD of 2.01 Å for modeling linker conformations in real-world conotoxin cases, outperforming Rosetta NGK (3.48 Å) in cases with significant conformational changes upon cyclization.
  • Linker Sequence Prediction: Experimental linker sequences were ranked significantly better (average percentile 37.2, p=0.024) than other potential sequences in the conotoxin benchmark.
  • Nrf2 Peptide Design: A cyclized Nrf2 peptide designed by PEP-Cyclizer showed a 26-fold increase in binding affinity to Keap1 compared to its linear counterpart.
  • Urotensin II Peptide Design: A head-to-tail cyclized Urotensin II analogue retained significant in vitro activity, demonstrating the method's ability to design functional cyclic peptides even for complex targets.

Limitations and Future Work

  • Linker Size Guessing: The current protocol requires the linker size to be specified, rather than automatically predicting it. This is identified as a subject for future investigation.
  • Predictive Power of Sequence Ranking: While experimental sequences were ranked better, the authors acknowledge that the limited number of positive cases and the lack of documented failures make it challenging to definitively assess predictive power.
  • Synthesis and Stability: PEP-Cyclizer predicts likely sequences and structures but does not predict the success of in vitro synthesis or the conformational stability of the resulting peptide.
  • Future Work: The authors suggest future research will focus on predicting the optimal linker length and further improving the predictive power of linker sequence prediction.

Summary updated

Abstract

Click to copy section linkSection link copied!
BibTeX
@article{karami2023exploring,
  author = {Yasaman Karami and Samuel Murail and Julien Giribaldi and Benjamin Lefranc and Florian Defontaine and Olivier Lesouhaitier and Jérôme Leprince and Sjoerd Jacob de Vries and Pierre Tufféry},
  title = {{Exploring a Structural Data Mining Approach to Design Linkers for Head-to-Tail Peptide Cyclization}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2023},
  url = {https://openreview.net/forum?id=f9rTo1X78z},
}
Figure