🧬 biology

Cyclome930: Large-scale replica-exchange dynamics of 930 cyclic peptide reveal thermal stability and critical metal-binding likelihood

This paper introduces Cyclome930, a comprehensive resource of 930 cyclic peptides featuring a novel symmetry-aware alignment algorithm and exhaustive molecular dynamics simulations, which collectively enable the development of machine learning models to predict thermal stability and critical metal-binding likelihood for computational peptide design.

Original authors: Ratul Chowdhury, Karuna Sajeevan, Hannah Gates, Vaishnavey, Curwen Tan, Riza Danurdoro, Julia Young

Published 2026-07-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ratul Chowdhury, Karuna Sajeevan, Hannah Gates, Vaishnavey, Curwen Tan, Riza Danurdoro, Julia Young

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine a world where tiny, circular necklaces made of amino acids—called cyclic peptides—are the ultimate superheroes of the molecular world. Unlike their floppy, straight-line cousins (linear peptides), these circular loops are tough, rigid, and incredibly hard to break. They are like the bodyguards of the protein world, resisting heat, enzymes, and chaos. But here's the problem: scientists have been trying to figure out exactly how strong these necklaces are and which ones are best at grabbing specific metals (like the rare minerals needed for electric cars) without a reliable map.

Enter Cyclome930, a massive new digital library created by researchers at Iowa State University and the Ames National Laboratory. Think of this library as a giant, organized attic where they've sorted through thousands of old boxes from different museums (databases like PDB, CyBase, and others) to find 930 unique, experimentally verified cyclic peptide structures. Before this, the best collections had structural info on only about 276 of them. This new library doesn't just list them; it gives them a home address, a family tree, and a detailed description of how they are tied together.

The "Rotating Door" Problem

One of the biggest headaches in studying these peptides is that they are circular. If you try to compare two circular necklaces by just looking at them as straight lines, you might miss the fact that they are actually twins. It's like trying to compare two identical bracelets by cutting them open at different spots; the patterns look totally different even though the beads are the same.

The paper argues that standard computer tools fail here. If you use a regular "linear" alignment method, you might think two peptides are only 50% similar when they are actually 100% identical, just rotated. The researchers built a new cyclicity-aware alignment tool to fix this. Instead of cutting the loop, their tool spins the necklace around, checking every possible starting point to find the best match. They found that this new method reveals hidden similarities that the old tools completely missed, especially for complex knots where the peptide is tied in multiple places.

The Heat Test: Simulating the Meltdown

How do you know if a peptide necklace will hold up in a hot oven? You can't just boil them all in a lab; it's too expensive and slow. So, the team ran 100-nanosecond replica-exchange molecular dynamics (REMD) simulations.

Imagine putting 930 different peptide necklaces into a virtual oven that slowly heats up from 298 K to 400 K (roughly room temperature to boiling water). They didn't just watch them; they tracked how the necklaces wiggled, stretched, and eventually started to unravel. They looked for a specific "melting point" (which they call STop2Melt) where the necklace suddenly goes from being a tight, compact ball to a loose, floppy mess.

For a famous peptide called Kalata B1, their simulation predicted a melting point of 382.38 K. This matched perfectly with real-world experiments that showed this peptide stays folded even when boiled. This gave them confidence that their virtual oven was accurate.

The Crystal Ball: Predicting Melting Points with AI

Now that they had the "melting points" for all 930 peptides from the simulations, they wanted to build a crystal ball (a machine learning model) to predict the melting point of any new peptide just by looking at its sequence.

They tried a few different approaches:

The "Linear" Approach: They fed the AI just the sequence of letters (amino acids) using a standard protein language model. Result: It failed miserably. The AI couldn't guess the melting point because it didn't understand the loop. It was like trying to guess how a bracelet holds up by only looking at a string of beads.
The "Cyclicity-Aware" Approach: They added a special "cyclic offset" feature that told the AI, "Hey, this bead is connected to that bead, even though they are far apart in the list." Result: The AI's accuracy jumped significantly.
The "Topological" Approach: They added even more details about the knot type (e.g., is it a simple loop, or does it have side-bridges?). Result: The model became even better, reaching a performance score (R²) of 0.76 with an error margin of about 7.2 K.

The paper explicitly states that ignoring the circular shape makes the model useless. You cannot predict how a loop behaves by treating it like a straight line.

The Metal Hunters: CritiCL

Finally, the researchers asked: "Which of these 930 tough necklaces are good at grabbing specific critical minerals like Cobalt (Co²⁺), Nickel (Ni²⁺), Manganese (Mn²⁺), or Lanthanides (Ln³⁺)?"

They trained a new classifier called CritiCL. This tool acts like a metal detector, scanning the 930 peptides to see which ones are most likely to bind to these specific metals. The model, which uses the same "cyclicity-aware" brain as the melting point predictor, successfully sorted the peptides into five different metal-binding categories. It found that many peptides showed high confidence in binding to specific metals, suggesting they could be used to pull these valuable minerals out of water or waste streams.

The Bottom Line

This paper doesn't claim to have solved the mystery of every peptide in the universe. Instead, it provides a foundational toolkit:

A database of 930 verified structures (Cyclome930).
A new way to compare circular sequences that doesn't get confused by rotation.
A simulation method to estimate how hot a peptide can get before melting.
An AI model (STop2Melt) that predicts this heat limit, but only if you tell it the peptide is a loop.
A screening tool (CritiCL) that suggests which loops might be good at catching critical metals.

The researchers emphasize that while their simulations are powerful, they are still simulations, not physical experiments for every single peptide. However, by combining physics-based simulations with machine learning that respects the circular nature of these molecules, they have built a much clearer path for designing future peptide-based tools for medicine and clean energy. The tools are now open-source, inviting others to play with the data and build even better necklaces.

Technical Summary: Cyclome930: Large-scale replica-exchange dynamics of 930 cyclic peptide reveal thermal stability and critical metal-binding likelihood

Problem Statement
Cyclic peptides offer significant advantages as therapeutic and industrial scaffolds due to their conformational rigidity, proteolytic resistance, and ability to bind targets with high specificity. However, their systematic analysis is hindered by fragmented data resources, a lack of standardized sequence comparison methods that account for rotational symmetry, and the absence of predictive frameworks for thermal stability and metal-binding propensity. Existing databases often lack comprehensive structural annotations or focus narrowly on specific families (e.g., cyclotides), while standard bioinformatics tools fail to recognize sequence equivalences in circular molecules, leading to inaccurate similarity scoring. Furthermore, there is a dearth of systematic data linking cyclic peptide topology to thermal resilience or their potential for critical mineral (CM) recovery.

Methodology
The authors present an integrated computational framework comprising four main components:

Database Curation (Cyclome930): The authors aggregated experimentally determined cyclic peptide structures from RCSB PDB, ConoServer, CyBase, and CyclicPepedia. After filtering for entries with 3D structures and removing sequence-level redundancy, they curated a non-redundant dataset of 930 unique cyclic peptides. This dataset includes metadata on source organisms, cyclization types, and structural coordinates, expanding the coverage of annotated cyclic peptides with structural data by approximately 3.4-fold compared to previous resources.
Cyclicity-Aware Sequence Alignment: To address the limitations of linear alignment, the authors developed a novel algorithm that accounts for rotational symmetry and knot topology. For end-to-end (e2e) cyclized peptides, the template sequence is duplicated and concatenated to form a "T-template," allowing the query to slide across all rotational windows to find the optimal alignment. For complex topologies (e.g., e2e+s2s), multiple T-templates are generated to reflect true covalent connectivity. This approach corrects the underestimation of similarity inherent in linear methods.
Physics-Based Stability Characterization (REMD): To generate training targets for thermal stability, the authors performed extensive all-atom replica-exchange molecular dynamics (REMD) simulations (100 ns per replica; 298–400 K) for the peptides in Cyclome930. They identified "melting-like" transitions ($STop2Melt$) by monitoring coordinated changes in backbone dihedral angles (Ramachandran maps), radius of gyration ( $R_g$ ), residue-wise root-mean-square fluctuation (RMSF), and root-mean-square deviation (RMSD). A transition was defined by a sharp expansion in $R_g$ (from <5 Å to >6 Å) and increased backbone flexibility.
Machine Learning Models:
- STop2Melt (Predictor): A supervised regression model was trained to predict $STop2Melt$ from sequence and topology. The model utilizes ESMc (Evolutionary Scale Modeling cyclic) embeddings augmented with "cyclic offset" features (encoding shortest-path distances under cyclic connectivity) and cyclization-class topology descriptors.
- CritiCL (Classifier): A multi-class classifier framework was developed to predict the metal-binding propensity of peptides toward critical minerals (Co²⁺, Ni²⁺, Mn²⁺, Ln³⁺, and others). This model was trained on linear protein/peptide data and applied in a zero-shot manner to the cyclic peptides in Cyclome930, incorporating cyclicity-aware embeddings.

Key Results

Dataset Expansion: Cyclome930 provides the largest curated resource of cyclic peptides with structural annotations to date, covering 244 source organisms and five distinct topological classes (e2e, s2e, s2s, e2e+s2s, s2e+s2s).
Alignment Accuracy: The cyclicity-aware alignment method demonstrated superior performance over linear methods. For e2e peptides, it recovered conserved motifs that linear alignment missed (e.g., increasing identity scores from 75% to 99% for specific pairs) and revealed asymmetric similarity matrices that reflect true topological relationships.
Thermal Stability Prediction: The STop2Melt model achieved strong predictive performance when incorporating cyclic topology. Models using only ESMc embeddings performed poorly ( $R^2 \leq 0.28$ ), whereas adding cyclic offsets and topology descriptors increased the test $R^2$ to 0.76 and reduced the Mean Absolute Error (MAE) to 7.2 K. Tree-based ensemble models (Random Forest, ExtraTrees) outperformed linear and neural network models, highlighting the non-linear relationship between topology and stability.
Validation: The REMD-derived $STop2Melt$ values were consistent with experimental observations, such as the high thermal stability of the cyclotide Kalata B1 (predicted ~382 K, consistent with experimental resilience up to ~370 K).
Metal Binding Screening: The CritiCL framework successfully classified peptides into five metal-binding categories. The model identified specific sequences with high confidence for binding critical minerals like Co²⁺ and Ln³⁺, suggesting potential for critical mineral recovery applications.

Significance and Claims
The paper claims to establish the first integrated, cyclicity-aware computational platform that combines database curation, topology-sensitive sequence analysis, exhaustive temperature-ramped molecular simulations, and machine learning. The authors assert that this work provides a foundational resource and cloud workbench (cyclome930.studio) for the computational design of cyclic peptide libraries.

Specifically, the study demonstrates that:

Standard linear bioinformatics tools are insufficient for cyclic peptides, and topology-aware methods are essential for accurate sequence analysis and machine learning.
Physics-based REMD simulations can generate reliable stability descriptors ($STop2Melt$) that serve as effective training targets for data-driven models.
Incorporating cyclic topology (via offsets and class descriptors) is a critical inductive bias for predicting thermal stability, significantly outperforming models that treat peptides as linear sequences.
The Cyclome930 resource and associated tools enable the systematic annotation of cyclic peptide libraries for biotechnology and critical mineral recovery, offering a pathway to engineer selective chelators for sustainable resource extraction.

The authors position this work as a step toward bridging the gap between small molecule drugs and protein therapeutics, and as a new lens for understanding the thermodynamic stability of cyclic polypeptides through the homological features of their spatial graphs.

The "Rotating Door" Problem

The Heat Test: Simulating the Meltdown

The Crystal Ball: Predicting Melting Points with AI

The Metal Hunters: CritiCL

The Bottom Line

More like this