Recognizing and generating knotted molecular structures by machine learning

Zhiyu Zhang #, Yongjian Zhu #, Liang Dai* · 2025

Read Recognizing and generating knotted molecular structures by machine learning by Zhiyu Zhang #, Yongjian Zhu #, Liang Dai* (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-12780.txt.

Introduction

Zhiyu Zhang, Yongjian Zhu, and Liang Dai teach neural networks to recognize and to generate knotted molecules. Knots appear in real molecules and are engineered for special properties, but telling one knot from another mathematically is slow, and designing a chain that ties a chosen knot was unsolved.

A transformer network classifies the knot type of a given chain shape with accuracy above ninety-nine percent. It handles many chain lengths in one model and runs about forty-five hundred times faster than the standard polynomial method. A second, diffusion-based network then dreams up new chains that tie a requested knot. Machine learning turns knot theory into a practical design tool.

Catalog record

CREATORZhiyu Zhang #, Yongjian Zhu #, Liang Dai*
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-12780

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Recognizing and generating knotted molecular structures by machine learning Zhiyu Zhang #, Yongjian Zhu #, Liang Dai* Abstract Knotted molecules occur naturally and are designed by scientists to gain special biological and material properties. Understanding and utilizing knotting require efficient methods to recognize and generate knotted structures, which are unsolved problems in mathematics and physics. Here, we solve these two problems using machine learning. First, our Transformer-based neural network (NN) can recognize the knot types of given chain conformations with an accuracy of > 99 % >99\% . We can use a single NN model to recognize knots with different chain lengths, and our computational speed is about 4500 times faster than the most popular mathematical method for knot recognition: the Alexander polynomials. Second, we for the first time design a diffusion-based NN model to generate conformations for given knot types. The generated conformations satisfy not only the desired knot types, but also the correct physical distributions of the radii of gyration and knot sizes. The results have several implications. First, the Transformer is suitable for handling knotting tasks, probably because of its strength in processing sequence information, a key component in knotting. Second, our NN can replace mathematical methods of knot recognition for faster speed on many occasions. Third, our models can facilitate the design of knotted protein structures. Lastly, analyzing how NN recognizes knot types can provide insight into the principle behind knots, an unsolved problem in mathematics. We provide an online website for using our models. Preprint: AIP/123-QED I Introduction Knotting is a common phenomenon in everyday objects and chain-like molecules, including DNA [ 1 , 2 , 3 , 4 , 5 ] and proteins [ 6 , 7 , 8 , 9 ] . It has attracted a broad range of scientists because knotting can impact many systems. In biology, DNA knots are involved in many key biological processes [ 10 , 11 , 12 , 13 , 14 ] , such as DNA replication, gene expression and cell division. DNA knots must be untied by topoisomerases in cells, otherwise, cells will die [ 15 ] . Hence, many cancer drugs kill cancer cells through inhibiting topoisomerase [ 16 ] . More than one thousand knotted protein structures have been discovered [ 17 , 9 ] . Why natural evolution chooses knotted proteins remains unclear. Some possible reasons are enhancing structural stability [ 18 , 19 ] and bringing special catalytic functions through knotting [ 20 ] . Similarly, chemists found that a manmade knotted structure has special catalytic functions [ 21 ] . In polymer physics, knotting can significantly slow down the relaxation of a compressed polymer [ 22 ] and the stretching kinetics of a polymer [ 23 ] . In addition, knotting can reduce the minimum breaking force of a polymer [ 24 ] . From the practical viewpoint, knotted structures have been suggested for controlled drug release because untying knots can release the substances wrapped by knots [ 25 ] . In nanotechnology, DNA knots occur in nanopores [ 5 , 26 , 27 , 28 ] and nanochannels [ 29 , 30 , 31 ] , which are designed for DNA analysis. Knotting may be utilized to achieve technical needs, such as slowing down nanopore translation for better data interpretation. Two foundations in the research of physical knotting are recognizing and generating knotted conformations. Recognizing a knotted conformation means identifying the knot type of a given conformation. This is a difficult task, in principle, not fully solved in math. Fig. 1 a shows a few conformations and their knot types. Recognizing their knot types by human eyes is impossible. Mathematicians have developed methods to recognize the knot types in the past century, including polynomial invariants (Alexander, Jones, and others) [ 32 , 33 , 34 ] , and homology invariants [ 35 , 36 ] . Knot invariants are defined based on the fact that for a closed rope, a quantity remains unchanged no matter how we distort the rope. This quantity is the invariant or knot type. This invariant is often a polynomial, e.g. t-1+t^{-1} for 3_{1} knot. A perfect knot-recognition method should accomplish two tasks: (i) for all conformations with the same knot type, the corresponding invariants (polynomials) are same; (ii) for conformations with different knot types, the corresponding polynomials are different. However, no method can accomplish the second tasks so far. For example, the method of the Alexander polynomials gives the same polynomial for the unknot and many other knot types. Recent studies have applied machine learning to recognize knot types for given conformations. In 2020, we used the Long Short-Term Memory (LSTM) model [ 37 ] and achieved an accuracy of 99% for polymer conformations with chain lengths of N=100 and five knot types: 0_{1} , 3_{1} , 4_{1} , 5_{1} , and 5_{2} [ 38 ] . The input feature is the bond vectors of polymer conformations. Later, Orlandini group [ 39 ] , Sulkowska group [ 40 ] , Michieletto group [ 41 ] , and Lu group [ 42 ] performed similar studies using different neural networks (NNs), different input features, and different chain lengths. The maximum chain length in most of these studies is N=200 . One practical application of these models is to hopefully replace the mathematical methods of knot recognition in molecular knot research, such as the popular Alexander polynomial method in recent studies [ 43 , 44 , 45 ] . NN models are often faster than mathematical methods in the knot recognition task. However, one limitation of all previous NN models is that they are trained for a fixed short chain length and can only be applied to that specific length due to RNN’s inability to learn long-term dependency [ 46 ] . This means that it is nearly impossible if one wants to use RNN for long polymer knot classification. To address the above chain length-generalization problem, we revolutionize the current RNN architecture with the Transformer [ 47 ] . Originally developed for Natural Language Processing, the Transformer model has demonstrated exceptional performance in tasks such as recognition [ 48 ] and text generation [ 49 , 50 ] . The reasoning behind our choice is the long-range learning enabled by the attention mechanism, which is paramount in capturing the topology of knots on the long polymer chain. Also, its parallel processing capability allows for fast computation, beating the current algorithm in calculating the Alexander polynomial. Using Transformer as our base architecture, we train a model for a wide range of chain lengths from 100 to 1000, which covers most chain lengths in the recent simulation studies of DNA knots [ 30 ] , polymer knots [ 43 , 44 , 45 ] , and most lengths of natural protein knots [ 17 , 9 ] . In addition, we expand the number of knot types to eight: 0_{1} , 3_{1} , 4_{1} , 5_{1} , 5_{2} , 6_{1} , 6_{2} , and 6_{3} , which also covers the most knot types in recent studies. Furthermore, our model achieves a higher accuracy than the LSTM models in previous studies. In addition to improving knot recognition, we for

Read the complete text