Cyclome: Large-scale replica-exchange dynamics of 930 cyclic peptide reveal thermal stability and critical metal-binding behavior

This study introduces Cyclome, a comprehensive computational framework that unifies a curated dataset of 930 cyclic peptides with novel topology-aware algorithms and machine learning models to predict thermal stability and identify critical metal-binding capabilities, thereby advancing the design of stable cyclic peptide therapeutics and tools for mineral recovery.

The Gist

Imagine you have a bunch of tiny, circular necklaces made of beads (amino acids). In the world of biology, these are called cyclic peptides. Unlike a standard necklace that has a clasp and a loose end, these are closed loops with no beginning or end. Scientists love them because they are tough, don't break down easily in the body, and can be shaped to grab onto specific targets, like a key fitting into a lock.

However, until now, scientists had a hard time studying them because:

1. The data was scattered across different libraries (like trying to find a specific book when the library is split into four different buildings).
2. Computers are bad at comparing circular things because they usually look at things in a straight line (like reading a sentence from left to right).
3. We didn't know which of these "necklaces" would stay strong if you heated them up, or which ones could grab onto rare metals (like those used in electric car batteries).

This paper, titled "Cyclome," is like building a massive, super-organized workshop to fix all these problems. Here is what they did, explained simply:

1. The Great Collection (Cyclome930)

Imagine trying to build a catalog of all the different types of circular necklaces in the world. Before this paper, the catalog was small and messy. The researchers went to four different digital "libraries" (databases) and gathered every single circular peptide they could find.

• The Result: They created Cyclome930, a master database of 930 unique, high-quality circular peptides. It's like taking 276 scattered puzzle pieces and finding 654 more to complete the picture, giving scientists a huge, clean dataset to work with.

2. The "Rotating" Translator (Cyclicity-Aware Alignment)

Here is the tricky part: If you have a circular necklace with the bead pattern "Red-Blue-Green," and someone else describes it as "Green-Red-Blue," a standard computer thinks they are different. But they are actually the same necklace, just rotated!

• The Problem: Old computer tools treat these like straight lines, so they miss the similarity.
• The Solution: The authors built a new "translator" algorithm. Imagine holding the necklace and spinning it around in your hand, checking every possible angle to see if it matches another one. This new tool realizes that "Red-Blue-Green" and "Green-Red-Blue" are actually the same thing, allowing for much more accurate comparisons.

3. The "Heat Test" (Simulating Melting)

Scientists wanted to know: If I heat this necklace up, when will it fall apart?

• The Experiment: They couldn't physically heat up 930 tiny necklaces in a lab (it would take forever and be expensive). Instead, they used a super-powerful computer simulation called REMD.
• The Analogy: Think of this like a video game where they put the necklaces in a virtual oven. They slowly turned up the heat from room temperature to boiling. They watched how the necklaces wiggled, stretched, and eventually unraveled.
• The Outcome: They figured out the exact "melting point" for each necklace. This tells us which ones are tough enough to survive in harsh environments (like inside a hot engine or a digestive system).

4. The Crystal Ball (Machine Learning)

Running those heat simulations takes a lot of computer power. So, the team used the results from the simulations to train a Machine Learning model (a type of AI) called STop2Melt.

• How it works: They taught the AI to look at the shape and pattern of the necklace and guess its melting point without needing to run the expensive heat simulation every time.
• The Secret Sauce: The AI was taught to understand the "circular" nature of the peptides. If you just gave it a straight-line description, it failed. But once they gave it a "circular map" (showing how the ends connect), it became a genius at predicting stability.

5. The Metal Catchers (CritiCL)

Finally, the team asked: Which of these necklaces are good at grabbing onto critical metals?

• Why it matters: We need to recover rare metals (like Cobalt, Nickel, and Lanthanides) from old electronics and batteries to build green technology.
• The Tool: They used another AI model called CritiCL. They fed the 930 necklaces into this model, and it predicted which ones would act like a magnet for specific metals.
• The Result: They created a "ranked list" of the best candidates. These are the peptides that could potentially be used to clean up electronic waste or mine metals from the ocean in a green, eco-friendly way.

The Big Picture

Think of this paper as the foundation for a new construction project.

• They built the library (Cyclome930).
• They invented a new language to describe the shapes (Cyclicity-aware alignment).
• They tested the strength of the materials (REMD simulations).
• They built a predictor to save time (STop2Melt).
• And they found the best tools for a specific job (CritiCL for metal mining).

By putting all these pieces together, they have given scientists a powerful toolkit to design better medicines and cleaner ways to recover the metals we need for our future technology.

Technical Summary

1. Problem Statement

Cyclic peptides are promising biomolecules for therapeutics and industrial applications due to their structural stability, resistance to degradation, and high binding specificity. However, their systematic analysis and computational design face three major bottlenecks:

• Fragmented Data: Existing databases (e.g., CyBase, ConoServer) are disjointed, lack comprehensive structural annotations, and cover only subsets of the cyclic peptide chemical space.
• Inadequate Sequence Analysis: Standard bioinformatics tools assume linear sequences with fixed N- and C-termini. Applying these to cyclic peptides fails to account for rotational symmetry and complex knot topologies, leading to inaccurate similarity scoring and motif identification.
• Lack of Stability Prediction: There is a scarcity of experimental thermal stability data (melting points) for cyclic peptides, and no established computational models exist to predict thermal stability based on sequence and topology. Furthermore, the potential of cyclic peptides for critical mineral (CM) recovery remains largely unexplored computationally.

2. Methodology

The authors developed an integrated computational framework comprising four main components:

A. Database Curation: Cyclome930

• Aggregation: Unified four public repositories (RCSB PDB, ConoServer, CyBase, CyclicPepedia) into a single resource.
• Filtering & Curation: Filtered for entries with experimentally determined 3D structures. After removing redundancy at the sequence level, they curated 930 unique cyclic peptides (Cyclome930), expanding the structural dataset by ~3.4-fold compared to previous efforts.
• Annotation: Annotated sequences with source organisms, cyclization types, and structural coordinates.
• Topological Classification: Classified peptides into five topological groups based on the First Betti number ($b_1(G)$) and linkage types:
  1. e2e: End-to-end (head-to-tail).
  2. s2e: Side-to-end.
  3. s2s: Side-to-side.
  4. e2e+s2s: End-to-end with side-to-side.
  5. s2e+s2s: Side-to-end with side-to-side.

B. Cyclicity-Aware Sequence Similarity

• Algorithm Development: Created a novel alignment algorithm that explicitly handles rotational symmetry.
  • For e2e peptides, the template sequence is duplicated and concatenated (T-template) to allow sliding window alignment across all possible rotational offsets.
  • For complex topologies (e2e+s2s, s2e+s2s), multiple T-templates are generated to reflect different traversal paths defined by the covalent bridges.
• Outcome: This approach corrects the underestimation of similarity found in linear alignments, revealing conserved motifs that are invisible to standard tools.

C. Physics-Based Simulation (REMD)

• Simulation: Performed extensive Replica-Exchange Molecular Dynamics (REMD) simulations (100 ns per replica) across a temperature range of 298 K – 400 K.
• Melting Point Extraction: Defined an inferred melting point (STop2Melt) based on structural transition indicators:
  • Sharp increase in Radius of Gyration ($R_g$) from <5 Å to >6 Å.
  • Broadening of Ramachandran dihedral angles ($\Delta\phi, \Delta\psi$).
  • Significant increases in Root-Mean-Square Fluctuation (RMSF) and Root-Mean-Square Deviation (RMSD).
• Validation: The simulated STop2Melt for Kalata B1 (382.38 K) aligned with experimental data confirming its stability up to boiling.

D. Machine Learning Models

• STop2Melt (Regression): A model to predict thermal stability from sequence and topology.
  • Features: Used ESMc (Evolutionary Scale Modeling) embeddings augmented with cyclic offsets (shortest-path distances in the cyclic graph) and topology descriptors (cyclization class).
  • Architecture: Evaluated nine regression models; tree-based ensembles (Random Forest, ExtraTrees) performed best.
• CritiCL (Classification): A multi-class classifier to predict critical mineral binding affinity.
  • Targets: Lanthanides ($Ln^{3+}$), Cobalt ($Co^{2+}$), Nickel ($Ni^{2+}$), Manganese ($Mn^{2+}$), and others.
  • Method: Zero-shot classification of Cyclome930 using cyclicity-aware embeddings.

3. Key Results

• Dataset Expansion: Cyclome930 is the largest curated cyclic peptide database with structural data (930 unique chains vs. ~276 in previous resources), covering 244 source organisms and diverse topological classes.
• Similarity Accuracy: The cyclicity-aware alignment revealed that linear methods significantly underestimate similarity. For example, a peptide pair with 75% linear similarity showed 99% cyclic similarity when rotational symmetry was accounted for. The method also correctly identified that linear scoring fails for complex multi-loop topologies (e2e+s2s).
• Thermal Stability Prediction:
  • Models using ESMc embeddings alone performed poorly ($R^2 \leq 0.28$), indicating linear sequence features are insufficient.
  • Adding cyclic offsets improved performance significantly ($R^2 \approx 0.66$).
  • The best model (STop2Melt) combined ESMc, cyclic offsets, and topology descriptors, achieving $R^2 = 0.76$ and MAE = 7.2 K. This confirms that topological constraints are critical determinants of thermal stability.
• Critical Mineral Binding:
  • The CritiCL model achieved high classification accuracy (XGBoost: 79% accuracy, 0.96 Macro ROC-AUC).
  • Screening Cyclome930 identified specific peptides with high predicted affinity for critical minerals ($Co^{2+}$, $Ln^{3+}$, etc.), providing a ranked library for potential use in clean energy technologies (e.g., EV battery recycling).

4. Significance and Impact

• Foundational Resource: Cyclome930 provides the first comprehensive, structurally annotated dataset for large-scale cyclic peptide analysis, enabling future research in design and discovery.
• Methodological Advance: The paper establishes that topology-aware representations are essential for both sequence analysis and machine learning in cyclic systems. It demonstrates that ignoring cyclic connectivity leads to fundamental errors in similarity scoring and stability prediction.
• Bridging Physics and AI: By using physics-based REMD simulations to generate training labels for ML models, the study creates a "physically grounded" predictor (STop2Melt) that can rapidly screen sequences without running expensive simulations.
• Sustainability Application: The identification of cyclic peptides with high affinity for critical minerals offers a novel, biologically inspired pathway for the sustainable recovery of rare earth elements and transition metals, addressing global supply chain challenges.
• Open Access: The authors have released the database, inference code, and web tools (cyclome930.studio) to the community, facilitating the design of stable cyclic peptide libraries.

In summary, this work transforms the study of cyclic peptides from a fragmented, linear-centric approach to a unified, topology-aware framework, enabling accurate prediction of thermal stability and the discovery of new materials for critical mineral recovery.