Method and device for constructing a hyperbolic discrete diffusion model of reverse folding of three-dimensional RNA structure

Abstract

The present application discloses a method for constructing a hyperbolic discrete diffusion model of reverse folding of a three-dimensional RNA structure. The method embeds the rich geometric features and other topological properties of the RNA three-dimensional structure into the hyperbolic space, uses the forward diffusion process to add stepwise noise to the real one-dimensional sequence of the RNA three-dimensional structure, and uses the hyperbolic equivariant graph neural network denoising model to perform stepwise denoising. On the basis of limited training samples, the nucleotide distribution of the target RNA three-dimensional structure, i.e., the one-dimensional sequence, is effectively restored. The method uses the hyperbolic equivariant graph neural network denoising model to represent the hierarchical structure of RNA and enhances the structural similarity measurement of three-dimensional reverse folding. The present invention provides a powerful artificial intelligence method for the generative design of functional RNA structures, which will contribute to the design of RNA drugs and RNA biomolecules and their application in biomedicine.

Description

Technical Field

[0001] The present application relates to the fields of biotechnology and biomedical technology, and in particular to a method and device for constructing a hyperbolic discrete diffusion model of reverse folding of a three-dimensional RNA structure (3D RNA). Background Art

[0002] Today, engineered functional ribonucleic acids (RNAs) are widely used in a variety of biomedical fields, including RNA-based gene editing, vaccines, therapeutics, and synthetic biology (Cao et al., 2024; Androsavich, 2024). Generative design algorithms provide unique opportunities for designing new RNA molecules with specific functions, and are expected to fully unleash the potential of RNA technology in a wide range of biomedical applications (Metkar, Pepin, and Moore 2024; Zhang et al. 2023; Leppek et al. 2022).

[0003] Similar to proteins, the 3D structure of RNA is an essential carrier of its multiple functions (Spitale and Incarnato 2023). Therefore, 3D structural reverse folding is a fundamental task for generative design of functional RNA, and the new one-dimensional (1D) RNA sequence generated can accurately fold into the desired 3D structure. Intuitively, 3D reverse folding is the inverse process of 3D structure prediction. However, in the context of RNA, 3D reverse folding faces unique challenges, and the main bottlenecks are reflected in three aspects: (1) The lack of high-quality RNA 3D structural data. Currently, there are less than 2000 high-quality RNA 3D structures determined experimentally recorded in the PDB database; (2) The complexity of the reverse folding problem: a given 1D sequence can fold into an optimal 3D structure, while for a given 3D structure, there may be many 1D sequences that can fold into the same 3D structural skeleton. In addition, this "one-to-many" design problem is further complicated by the hierarchical folding constraints of RNA; (3) Efficient representation learning of RNA 3D structure: On the one hand, representation learning methods that can sensitively capture structural differences are crucial for 3D RNA reverse folding, because small changes in RNA 3D structure may lead to significant changes or even loss of function. On the other hand, the “low-sample” challenge further highlights the importance of representation learning efficiency for effective model training. Therefore, efficient 3D RNA inverse folding tools are urgently needed to address these challenges.

[0004] Well-folded RNA molecules exhibit a wide range of biological functions by interacting with other molecules such as small chemicals, DNA, and proteins, which is often driven by their three-dimensional structure (Corley, Burns, and Yeo, 2020; Spitale and Incarnato, 2023; Childs-Disney et al. 2022). RNA three-dimensional structure is a hierarchical geometric structure, including one-dimensional (1D) structure, secondary (2D) structural motifs, and long-range interactions between different secondary motifs. RNA 1D structure is a one-dimensional sequence composed of four nucleotides, which can be classified according to their nitrogenous bases: adenine (A), cytosine (C), guanine (G), and uracil (U). The 1D structure of RNA first folds into a secondary structure according to base pairing rules (AU, GC, or GU), and then forms a complex three-dimensional structure. However, base pairs may form between different secondary structural motifs, which are called pseudoknots and are common three-dimensional structural features in functional RNAs, such as ribozymes. The nucleotides in the pseudoknot are adjacent in 3D space but far apart in the 1D sequence. Therefore, for long RNA sequences, there are many possible local and long-range base pair combinations during the folding process, and the mechanistic mechanisms are not fully understood. Therefore, hierarchical geometric 3D global and local features as well as local and long-range base pairing constraints in 1D and 3D views must be jointly considered for effective sample representation learning for the “one-to-many” 3D RNA inverse folding problem.

[0005] This urgent need prompted us to develop an efficient 3D RNA inverse folding model to infer the optimal one-dimensional sequence given a three-dimensional RNA structural backbone. Although some methods have explored the inverse folding problem, existing inverse folding models have obvious limitations: (1) They mainly use local updated autoregressive generation methods, which fail to fully consider the overall structure and functional requirements of RNA sequence design. (2) Most models only use local information as features, ignoring the importance of long-range interactions and global structures in 3D space.

[0006] RNA structure prediction has been a challenging and evolving field in computational biology, with significant progress achieved through advances in various methods. Early RNA secondary structure prediction methods mainly relied on thermodynamic models and dynamic programming algorithms to predict RNA secondary structures using minimum free energy. Tools such as MFold (Zuker 2003), RNAfold (Hofacker 2003; Lorenz et al. 2011), and RNAstructure (Reuter and Mathews 2010) implemented these methods. After that, researchers began to use machine learning techniques to improve prediction accuracy. Tools such as CONTRAfold (Do, Woods, and Batzoglou 2006) and ContextFold (Zakov et al. 2011) adopted statistical model-based methods to learn RNA secondary structure features from a large amount of experimental data. After that, researchers began to use machine learning techniques to improve the accuracy of three-dimensional structure prediction. Tools such as RNAMake (Yesselman et al. 2011) have been widely used in the field of RNA structure prediction.

[0007] et al. 2019) and FARFAR2 (Watkins, Rangan and Das 2020) use statistical model-based methods to learn RNA three-dimensional structural features using a large amount of experimental data. In recent years, deep learning techniques have also been widely used in RNA three-dimensional structure prediction. Tools such as Rhofold (Shenet al. 2022), trRosettaRNA (Wang et al. 2023) and (Li et al. 2023) use neural network models to directly predict the 3D structure of RNA, using end-to-end learning methods to learn the process of folding from one-dimensional sequences to 3D structures. Recently, some studies have attempted to solve the RNA reverse folding task and partially addressed its challenges. However, they have not yet fully utilized global geometric information to capture the complex relationship between 3D structure and 1D sequence.

[0008] RNA reverse folding is an important task in the design of functional RNA generation, which brings revolutionary opportunities for various RNA-based biotechnology and biomedical applications. Specifically, RNA reverse folding refers to the generation of new RNA sequences that can fold into target topological structures. However, existing RNA reverse folding algorithms are mainly developed for secondary structure design. In three-dimensional (3D) RNA design, RNA reverse folding still faces great challenges due to the limited experimentally derived 3D structural data and the unique characteristics of RNA 3D structure.

[0009] Therefore, there is an urgent need in this field to develop a method and device for constructing a hyperbolic discrete diffusion model of 3D RNA reverse folding, by embedding the rich geometric features and other topological properties of the RNA three-dimensional structure into the hyperbolic space, and using the discrete diffusion model to infer the optimal one-dimensional sequence of a given three-dimensional RNA structure skeleton. Summary of the invention

[0010] The purpose of the present application is to provide a method and device for constructing a hyperbolic discrete diffusion model of the reverse folding of a three-dimensional RNA structure, by embedding the rich geometric features and other topological properties of the RNA three-dimensional structure into the hyperbolic space, and using the discrete diffusion model to infer the optimal one-dimensional sequence of a given three-dimensional RNA structure skeleton.

[0011] The first aspect of the present application provides a method for constructing a hyperbolic discrete diffusion model of reverse folding of a three-dimensional RNA structure, comprising the following steps:

[0012] (a) obtaining a three-dimensional structure of RNA and its true one-dimensional sequence, and abstractly representing the three-dimensional structure of RNA through a graph structure, wherein the graph structure includes node features and edge features of the three-dimensional structure of RNA;

[0013] (b) performing a forward diffusion process on the real one-dimensional sequence of the RNA three-dimensional structure, in which noise is iteratively added to the real one-dimensional sequence of the RNA three-dimensional structure until the probability distribution of the nucleotide sequence AUGC of the real one-dimensional sequence is uniformized, and outputting a noisy nucleotide sequence with a uniform probability distribution, wherein when noise is added at each time step, the transition probability at each time step is calculated;

[0014] (c) randomly sampling from the noise nucleotide sequence to obtain a random noise nucleotide sequence;

[0015] (d) inputting the edge features of the RNA three-dimensional structure into a multi-layer perceptron to obtain edge embedding;

[0016] (e) obtaining a hyperbolic embedding of a combination of node features including the node features of the RNA three-dimensional structure and the node features of the random noise nucleotide sequence through a hyperbolic space model;

[0017] (f) inputting the hyperbolic embedding of the combined node feature and the edge embedding into a hyperbolic equivariant graph neural denoising network model, training the hyperbolic equivariant graph neural network denoising model, updating the combined node feature in the hyperbolic equivariant graph neural denoising network model, and obtaining a predicted probability distribution of the intermediate nucleotide sequence based on the updated combined node feature;

[0018] Based on the obtained predicted probability distribution of the intermediate nucleotide sequence and the transition probability at the corresponding time step obtained in step (b), the conditional probability distribution of the intermediate nucleotide sequence is obtained, thereby completing a denoising process;

[0019] (g) adopting an intermediate one-dimensional RNA sequence from the probability distribution of the intermediate nucleotide sequence;

[0020] (h) Iterating steps (e)-(g) multiple times, during the iteration, replacing the random noise nucleotide sequence in step (e) with the intermediate RNA one-dimensional sequence obtained in step (g), and after the iteration is completed, obtaining the optimal RNA one-dimensional sequence by sampling from the probability distribution of the denoised nucleotide sequence.

[0021] In another preferred embodiment, step (a) also includes extracting the RNA three-dimensional skeleton structure of the RNA three-dimensional structure, wherein the RNA three-dimensional skeleton structure includes RNA skeleton atoms, and the RNA skeleton atoms include three atoms in each nucleotide, and each nucleotide is represented by the three atoms: P, C4', and N1 or N9, wherein N1 is A or G, and N9 is C or U. Node features and edge features are constructed by the three skeleton atoms and their relationships, and the nucleotide node is represented as the C4' atom, and the edge is determined by the K proximity graph, or when the nucleotide is A or G, the third skeleton atom is the atom at the N1 position, and when the nucleotide is C or U, the third skeleton atom is the atom at the N9 position.

[0022] In another preferred embodiment, a k-neighbor graph G(X, E) is used to represent the RNA three-dimensional skeleton structure, where E is a set of edge features of the RNA three-dimensional skeleton structure, and X is a set of node features of the RNA three-dimensional skeleton structure.

[0023] In another preferred embodiment, in the RNA three-dimensional skeleton structure, each nucleotide is represented by three skeleton atoms, and the node features and edge features of the RNA three-dimensional skeleton structure are constructed by the three skeleton atoms and their relationships.

[0024] In another preferred embodiment, RNA is represented by a set of nodes, each node is a nucleotide residue in RNA, and each nucleotide node i has a three-dimensional coordinate , which are the three-dimensional coordinates of residue C4'.

[0025] In another preferred embodiment, the node characteristics of the RNA three-dimensional skeleton include nucleotide type, physicochemical properties, local environmental information, and the position of the nucleotide node, and the position of the nucleotide node is the three-dimensional coordinate of the nucleotide node.

[0026] In another preferred example, the node characteristics of the random noise nucleotide sequence include nucleotide types.

[0027] In another preferred example, the combined node feature is formed by splicing the node feature of the random noise nucleotide sequence and the node feature of the random noise nucleotide sequence.

[0028] In another preferred embodiment, the hyperbolic space model in the step (e) is a Poincare disk.

[0029] In another preferred embodiment, in the step (b), uniformly distributed noise is iteratively added to the three-dimensional RNA skeleton structure.

[0030] In another preferred example, in the step (f), the hyperbolic embedding of the combined node feature is updated to obtain an updated hyperbolic embedding of the combined node feature, and the updated hyperbolic embedding of the combined node feature is mapped into an updated Euclidean encoding of the combined node feature, and the predicted probability distribution of the intermediate nucleotide sequence is obtained based on the updated Euclidean encoding of the combined node feature.

[0031] In another preferred embodiment, in the step (e), the nucleotide node features of the RNA three-dimensional skeleton structure and the node features of the random nucleotide sequence are spliced ​​together and then input into the hyperbolic space model to obtain the hyperbolic embedding of the combined nucleotide node features.

[0032] In another preferred example, in the step (e), the nucleotide node features of the RNA three-dimensional skeleton structure and the node features of the random nucleotide sequence are respectively input into the hyperbolic space model, and the obtained outputs are spliced ​​together to obtain the hyperbolic embedding of the combined nucleotide node features.

[0033] In another preferred example, the hyperbolic equivariant denoising network model includes one or more layers of hyperbolic equivariant graph convolution layers.

[0034] In another preferred example, in the training of the hyperbolic equivariant denoising network model, the update of all node features is the update of the combined node features.

[0035] In another preferred embodiment, in step (f), during the training of the hyperbolic equivariant graph neural network model, the edge embedding of the input Sub-step to perform update (f1):

[0036] First, use a logarithmic mapping , the hidden representation of the combined node Projection to tangent space , get the combined node in Euclidean encoding of features in layer networks ;

[0037] Then, the following formula is used to obtain +1 layer of updated edge embedding :

[0038] Encode distances and current edge embeddings using hyperbolic space get,

[0039] The updated edge embedding The definition is as follows:

[0040] ,

[0041] in, is a learnable function parameterized by a neural network, i and j represent the index of the combined node, where j is the neighbor node of i, Represents the network layer, h refers to the node hidden layer embedding, Indicates that the combination node i is in Euclidean coding in layer networks, It means that the combination node j is in the Euclidean encoding in layer networks, d D It represents the relative distance between the combined nodes i and j in the hyperbolic space, E represents the Euclidean space, and D represents the hyperbolic space.

[0042] In another preferred embodiment, the combined node feature includes the position of the nucleotide node in the node feature of the RNA three-dimensional skeleton;

[0043] In step (f), during the training process of the hyperbolic equivariant graph neural network model, a sub-step (f2) is further included: updating the hyperbolic position of the nucleotide node, encoding the updated hyperbolic position of the combined node feature Based on the current Euclidean node encoding , the updated hyperbolic position encoding Updated by the following formula:

[0044] ,

[0045] ,

[0046] in is a learnable function parameterized by a neural network, t represents the time step, σ is the activation function, exp c is the exponential transformation operation;

[0047] Step (f3), the updated hyperbolic embedding of the combined node features is obtained by the following formula :

[0048] ,

[0049] ,

[0050] in, is a learnable function parameterized by a neural network.

[0051] In another preferred embodiment, the number of iterations mentioned in step (h) may be 500.

[0052] The second aspect of the present application provides a method for predicting a one-dimensional RNA sequence based on a given three-dimensional RNA structure, comprising the following steps:

[0053] (a) obtaining the three-dimensional structure of RNA and its true one-dimensional sequence, and extracting the RNA three-dimensional skeleton structure of the RNA three-dimensional structure, and using a k-neighbor graph G(X,E) to represent the RNA three-dimensional skeleton structure, where E is a set of edge features of the RNA three-dimensional skeleton structure, and X is a set of node features of the RNA three-dimensional skeleton structure;

[0054] (b) performing a forward diffusion process on the real one-dimensional sequence of the RNA three-dimensional structure, during which noise is iteratively added to the real one-dimensional sequence of the RNA three-dimensional skeleton structure until the probability distribution of the nucleotide sequence AUGC of the real one-dimensional sequence is uniformized, and outputting a noisy nucleotide sequence with a uniform probability distribution, wherein when noise is added at each time step, the transition probability at each time step is calculated;

[0055] (c) randomly sampling from the noise nucleotide sequence to obtain a random noise nucleotide sequence;

[0056] (d) inputting the edge features of the RNA three-dimensional skeleton structure into a multi-layer perceptron to obtain edge embedding;

[0057] (e) obtaining a hyperbolic embedding of a combination of node features including the node features of the RNA three-dimensional skeleton structure and the node features of the random noise nucleotide sequence through a hyperbolic space model;

[0058] (f) during the denoising process, the hyperbolic embedding of the combined node features and the edge embedding are input into the trained hyperbolic equivariant graph neural denoising network model, the combined node features are updated in the hyperbolic equivariant graph neural denoising network model, and the predicted probability distribution of the intermediate nucleotide sequence is obtained based on the updated combined node features;

[0059] Based on the obtained predicted probability distribution of the intermediate nucleotide sequence and the transition probability at the corresponding time step obtained in step (b), the probability distribution of the intermediate nucleotide sequence is obtained, thereby completing a denoising process;

[0060] (g) adopting an intermediate one-dimensional RNA sequence from the probability distribution of the intermediate nucleotide sequence;

[0061] (h) Iterating steps (e)-(g), replacing the random noise nucleotide sequence in step (e) with the intermediate RNA one-dimensional sequence obtained in step (g), thereby obtaining an optimal RNA one-dimensional sequence.

[0062] The third aspect of the present application provides a device for predicting a one-dimensional RNA sequence based on a hyperbolic discrete diffusion model, comprising:

[0063] a memory for storing computer executable instructions; and,

[0064] A processor, coupled to the memory, is configured to implement the steps in the above method when executing the computer executable instructions.

[0065] A fourth aspect of the present application provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps in the method described above are implemented.

[0066] A fifth aspect of the present application provides a computer program product, comprising computer executable instructions, which implement the steps in the above method when executed by a processor.

[0067] It should be understood that within the scope of the present invention, the above-mentioned technical features of the present invention and the technical features specifically described below (such as embodiments) can be combined with each other to form a new or preferred technical solution. Due to space limitations, they will not be described one by one here.

[0068] This application has at least one of the following advantages

[0069] (a) The hyperbolic discrete diffusion model of 3D RNA reverse folding of the present application uses a hyperbolic equivariant graph neural network denoising model to represent the hierarchical structure of RNA and enhance the structural similarity metric of 3D reverse folding;

[0070] (b) The present application uses a residue characterization method to construct complex residue features through local and long-range interactions. In the present application, residues refer to nucleotide residues and nucleotide nodes;

[0071] (c) This application conducted a large number of experiments on different datasets to evaluate the generalization ability of the model. The experimental results show that the accuracy and efficiency of this method are better than the existing methods, verifying the effectiveness of this method.

[0072] (d) The present invention provides a powerful artificial intelligence method for the generative design of functional RNA structures, which will facilitate the design of RNA drugs and RNA biomolecules and their application in biomedicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. It should be understood that the drawings described below are only some implementation examples of the present invention, and ordinary technicians in this field can also obtain other implementation examples based on these drawings without paying creative work.

[0074] Figure 1 is the embedding of the RNA molecule backbone in Euclidean space and its embedding in hyperbolic space (Poincare disk), where, as shown in the figure, the distances grow faster in the Poincare disk than in the Euclidean space, thus better distinguishing structural similarities between molecules;

[0075] Figure 2 It is a schematic diagram of the training process of the hyperbolic discrete diffusion model of 3D RNA inverse folding of the present application, wherein the hyperbolic discrete diffusion model (RI) diffusion overview: during the diffusion process, uniformly distributed noise is gradually added to the nucleotide sequence of the real RNA molecule, causing the structure represented by the sequence to undergo random transformations until the sequence distribution is completely transformed into uniformly distributed noise; during the denoising process, the nucleotide sequence is first randomly sampled from the uniform distribution, and then gradually denoised by a denoising neural network, and finally a real nucleotide sequence is generated;

[0076] Figure 3 is a schematic diagram of an improved Transformer module for encoding the position information of nucleotide node features;

[0077] Figure 4 These are some results of predicting one-dimensional sequence based on a given RNA three-dimensional structure using the hyperbolic discrete diffusion model of 3D RNA inverse folding of this application. DETAILED DESCRIPTION

[0078] Through extensive and in-depth research, the inventors have developed for the first time a hyperbolic discrete diffusion model RIdiffusion (including a hyperbolic equivariant graph neural network denoising model) for 3D RNA reverse folding for RNA reverse folding in 3D RNA design. This method embeds the rich geometric features and other topological properties of the RNA three-dimensional structure into the hyperbolic space, and uses the discrete diffusion model to effectively restore the nucleotide distribution (i.e., one-dimensional sequence) of the target RNA three-dimensional structure based on limited training samples. RIdiffusion was extensively evaluated using different data sets and strict data segmentation strategies, and the results showed that RIdiffusion consistently outperformed the baseline and the latest generative models in RNA reverse folding. This study provides a powerful tool for the generative design of functional RNA, which will be beneficial to a variety of downstream applications.

[0079] In the following description, many technical details are provided to help readers better understand the present application. However, those skilled in the art can understand that the technical solution claimed in the present application can be implemented even without these technical details and various changes and modifications based on the following embodiments.

[0080] To make the purpose, technical solution and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below in conjunction with the accompanying drawings. It should be understood that these are only some examples that the present invention may take for the reader, but are not intended to limit the scope of the present invention.

[0081] the term

[0082] As used herein, "nucleotide residue" and "nucleotide" are used interchangeably.

[0083] The three-dimensional backbone of RNA

[0084] First, we extract the skeleton atoms from the original RNA structure and use them to construct the original features. We regard the nucleotides as a graph data structure G=(X,E). G is the graph, X is the set of nucleotide nodes, and RNA is represented by a set of nodes, each of which corresponds to a nucleotide residue in RNA. Each nucleotide node i has a three-dimensional coordinate , which is the three-dimensional coordinate of residue C4'. E is the edge set. The local coordinate system can be established accordingly. There are node features and edge features correspondingly.

[0085] Nodal features of RNA three-dimensional backbone structure

[0086] Includes: 1) Dihedral angle: calculated by the angle formed between three backbone atoms; 2) Nucleotide type: four-dimensional one-hot encoding to represent AUGC; 3) Surface perception features: 4) Position information: three-dimensional coordinates of backbone atoms (nodes); 5) Physical and chemical properties: SASA and b-factor values.

[0087] Edge features of RNA three-dimensional backbone structure

[0088] Including: 1) Edge features based on Gaussian kernel distance: For two nucleotide nodes x i With x j , this feature between them is obtained through the RBF function: RBF ||x i -x j ||.

[0089] 2) Relative edge feature: The vector form of the difference between the absolute positions of two nucleotide nodes: the absolute position difference of the three-dimensional coordinates of the nodes (x i -x j ) is obtained by multiplying the basis of the local coordinate system.

[0090] 3) Relative edge direction feature: After establishing the local coordinate system, the direction vector of the nucleotide node is obtained by multiplying the basis of the local coordinate system.

[0091] 4) Binary contact signal: Determine whether there is contact between two nucleotide nodes based on the cutoff distance (0 for no contact and 1 for contact).

[0092] It should be noted that in the application documents of this patent, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one" do not exclude the existence of other identical elements in the process, method, article or device including the elements. In the application documents of this patent, if it is mentioned that an action is performed according to an element, it means that the action is performed at least according to the element, which includes two situations: performing the action only according to the element, and performing the action according to the element and other elements. Expressions such as multiple, multiple, and multiple include 2, 2 times, 2 kinds, and more than 2, more than 2 times, and more than 2 kinds.

[0093] A method for constructing a 3D RNA inverse folding hyperbolic discrete diffusion model (RIdiffusion)

[0094] The first aspect of the present application provides a method for constructing a hyperbolic discrete diffusion model of 3D RNA inverse folding, comprising the following steps:

[0095] (a) obtaining a three-dimensional structure of RNA and its true one-dimensional sequence, and abstractly representing the three-dimensional structure of RNA through a graph structure, wherein the graph structure includes node features and edge features of the three-dimensional structure of RNA;

[0096] (b) performing a forward diffusion process on the real one-dimensional sequence of the RNA three-dimensional structure, in which noise is iteratively added to the real one-dimensional sequence of the RNA three-dimensional skeleton structure until the probability distribution of the nucleotide sequence AUGC of the real one-dimensional sequence is uniformized, and outputting a noisy nucleotide sequence with a uniform probability distribution, wherein when noise is added at each time step, the transition probability at each time step is calculated;

[0097] (c) randomly sampling from the noise nucleotide sequence to obtain a random noise nucleotide sequence;

[0098] (d) inputting the edge features of the RNA three-dimensional skeleton structure into a multi-layer perceptron to obtain edge embedding;

[0099] (e) obtaining a hyperbolic embedding of a combination of node features including the node features of the RNA three-dimensional skeleton structure and the node features of the random noise nucleotide sequence through a hyperbolic space model;

[0100] (f) inputting the hyperbolic embedding of the combined node feature and the edge embedding into a hyperbolic equivariant graph neural denoising network model, training the hyperbolic equivariant graph neural network denoising model, updating the combined node feature in the hyperbolic equivariant graph neural denoising network model, and obtaining a predicted probability distribution of the intermediate nucleotide sequence based on the updated combined node feature;

[0101] Based on the obtained predicted probability distribution of the intermediate nucleotide sequence and the transition probability at the corresponding time step obtained in step (b), the conditional probability distribution of the intermediate nucleotide sequence is obtained, thereby completing a denoising process;

[0102] (g) adopting an intermediate one-dimensional RNA sequence from the probability distribution of the intermediate nucleotide sequence;

[0103] (h) Iterating steps (e)-(g) multiple times, during the iteration, replacing the random noise nucleotide sequence in step (e) with the intermediate RNA one-dimensional sequence obtained in step (g), and after the iteration is completed, obtaining the optimal RNA one-dimensional sequence by sampling from the probability distribution of the denoised nucleotide sequence.

[0104] Preferably, in step (b), the transition probability is obtained by the following formula:

[0105] .

[0106] Preferably, in step (f), the probability distribution of the intermediate nucleotide sequence based on the updated combined node features is obtained by the following formula: :

[0107] .

[0108] in, is the predicted probability distribution of the nucleotide sequence, which is predicted by the hyperbolic equivariant graph neural network denoising model.

[0109] Preferably, in the denoising process, noise is added at each time step t, and the denoising is performed at t=0, t=1, ... t=T-1, t=T times, and the denoising is completed at t=T; then a denoising process is performed, and denoising is performed at t=T, t=T-1, ... t=1, t=0 times. Specifically, the predicted probability distribution predicted in the denoising process from t=T to t=T-1 and the posterior distribution calculated from the transfer matrix obtained in the denoising process at t=0, t=1 are used together to obtain the conditional probability distribution of the intermediate nucleotide sequence.

[0110] Preferably, step (a) also includes extracting the RNA three-dimensional skeleton structure of the RNA three-dimensional structure, wherein the RNA three-dimensional skeleton structure includes RNA skeleton atoms, and the RNA skeleton atoms include three atoms in each nucleotide, and each nucleotide is represented by the three atoms: P, C4', and N1 or N9, wherein N1 is A or G, and N9 is C or U. Node features and edge features are constructed by the three skeleton atoms and their relationships, and the nucleotide node is represented as the C4' atom, and the edge is determined by the K proximity graph, or in other words, when the nucleotide is A or G, the third skeleton atom is the atom at the N1 position, and when the nucleotide is C or U, the third skeleton atom is the atom at the N9 position.

[0111] Preferably, in step (a), it also includes using a k-neighbor graph G(X, E) to represent the RNA three-dimensional skeleton structure, wherein E is a set of edge features of the RNA three-dimensional skeleton structure, and X is a set of node features of the RNA three-dimensional skeleton structure.

[0112] Hyperbolic discrete diffusion model for 3D RNA reverse folding

[0113] This application develops an efficient 3D RNA inverse folding model to infer the optimal one-dimensional sequence for a given three-dimensional RNA structural backbone.

[0114] This application proposes a 3D RNA inverse folding problem with a hyperbolic discrete diffusion model, defining it as a denoising problem. Discrete diffusion models operate in a discrete state space, iteratively and independently predicting each residue type and gradually generating finer structures, thereby capturing the global complex relationship between 1D nucleotide sequence and 3D structure, as well as the diversity of nucleotide sequences in RNA molecules. This makes them particularly suitable for sequence generation tasks. In order to capture information about long-range spatial interactions, this application employs Transformer layers and encoded spatial locations. In addition, this application recognizes that RNA molecules (secondary structure elements such as stem-loops, pseudoknots, etc.) naturally have a hierarchical structure that can be more efficiently represented in hyperbolic space. Compared to Euclidean space, hyperbolic distances better capture the differences in molecules, which is crucial for representing subtle changes between structures, such as Figure 1 As shown. The diffusion model is parameterized using a hyperbolic equivariant graph neural network (HEGNN), and the geometric information and topological properties of RNA molecules are incorporated into the model generation process, thereby more effectively capturing the three-dimensional structural characteristics of RNA molecules. In order to comprehensively evaluate the effectiveness of the model, this application constructs a series of datasets by dividing the RNAsolo dataset (a mature benchmark dataset) based on different sequence lengths and sequence similarity thresholds. Experimental results show that on datasets with low sequence similarity, the model has a 7.27% advantage over the suboptimal model, and its performance in multiple benchmarks is consistently better than the baseline and the latest (State-Of-The-Art, SOTA) models.

[0115] 3D RNA reverse folding problem

[0116] Given an RNA sequence , where S i Represents the i-th nucleotide, and the conformation C of the RNA molecule in 3D space. The traditional folding problem involves predicting the most likely structure C from the sequence S. In contrast, the inverse folding problem aims to deduce a suitable sequence S from the target structure C. The naturally occurring RNA sequence consists of four nucleotides. Formally, the RNA inverse folding problem can be expressed as the following optimization problem:

[0117]

[0118] Discrete Diffusion Model

[0119] Discrete diffusion models are a type of diffusion model that is good at processing discrete data (such as text data such as nucleotide sequences). They better capture the structural information of discrete data by defining more structured discrete destruction processes, such as simulating Gaussian kernels and nearest neighbor-based embedding spaces. The hyperbolic discrete diffusion model of 3D RNA inverse folding in this application includes a hyperbolic equivariant graph neural network denoising model.

[0120] Diffusion process

[0121] In the forward process, noise is continuously added to the data until it is transformed into pure noise. This process can be represented by a transition matrix, which facilitates the transition from the current time step to the next time step. The transition matrix is ​​iteratively applied to the original data, which evolves over time due to the influence of the noise. As the number of time steps increases, the original data will eventually converge to a uniform distribution for all data types. For a discrete variable with n types, the transition probabilities can be represented by the matrix [ Q t ] ij = q ( x t = j ∣ x t - 1 = i ) Based on Q t , you can set the following transformation kernels:

[0122] and (2)

[0123] in It's Q t Cumulative version of .

[0124] Parameterization of the inverse process

[0125] The inverse process aims to recover the original data distribution by removing the noise step by step. This process iteratively refines the noisy data back to its original form. The reverse process is guided by a learned denoising function that predicts the denoised data at each step and aims to approximate the inverse of the forward process. New samples can be iteratively generated by back diffusion. Probability distribution is the probability predicted by the hyperbolic equivariant graph neural network denoising model Obtained. By marginalizing the network predictions the following parameterization is obtained:

[0126]

[0127] The latter can be written as:

[0128]

[0129] in yes The classification distribution of .

[0130] Poincare sphere model

[0131] Riemannian geometry

[0132] Riemannian geometry is a powerful branch of mathematics that studies the properties of manifolds with a Riemannian metric. It is widely used to solve non-Euclidean geometry problems in machine learning. A Riemannian manifold is a smooth manifold with a Riemannian metric, denoted by At every point , there is a cutting space . Riemann metric is the inner product function defined on these tangent spaces The Riemann metric is used to measure distances on a manifold.

[0133] Hyperbolic space refers to a Riemann manifold with constant negative curvature. Poincare disk-sphere is one of the fundamental mathematical models representing hyperbolic geometry (Cannon et al. 1997). It characterizes the properties of hyperbolic geometry by mapping the entire hyperbolic plane onto the interior of a disk, which consists of the manifold Definition, the Riemann metric is: and the radius is The Poincare sphere.

[0134] distance

[0135] For a given point , the induced distance can be written as:

[0136]

[0137] Exponential and Logarithmic Mappings

[0138] In hyperbolic space, exponential and logarithmic maps are mathematical tools used to convert points between hyperbolic space and its tangent space. These maps facilitate geometric operations between hyperbolic and Euclidean space.

[0139] for , exponential mapping : The point in the hyperbolic space Assign to tangent vector , is given by:

[0140]

[0141] The logarithmic map is the inverse of the exponential map. It maps a point x in the hyperbolic space back to a vector v in the tangent space. The logarithmic map is given by:

[0142]

[0143] in, It is Müobius addition.

[0144] Parallel movement

[0145] In hyperbolic geometry, parallel transfer is an operation that moves a vector from one point to another while preserving the vector's angle and length relative to its starting point. Specifically, parallel transfer is a transformation that preserves the Riemann metric tensor, including its norm and inner product. Given a smooth manifold M, parallel transfer Vector Mapping .

[0146] RIdiffusion: a hyperbolic discrete diffusion model for 3D RNA reverse folding

[0147] The denoising network largely determines the quality of the generated diffusion model. In this section, a hyperbolic denoising network RI diffusion is proposed, which embeds molecular structural features into an equivariant hyperbolic space to capture the molecular hierarchy and measure the similarity between molecules, such as Figure 2 shown.

[0148] Residue map of folded RNA

[0149] The geometric structure of RNA is crucial to its function. In order to obtain the complex geometric information of folded RNA molecules, a k-nearest neighbor graph G = (X, E) is used to represent the three-dimensional structure of RNA, which describes the physicochemical properties and geometric information of RNA molecules.

[0150] Edge feature E = [E rbf , E lf , E cont , E sr ] contains the relationships between connected nodes, which are determined by the encoding scheme, such as relative spatial distance and sequence distance. For adjacent residues iii and jjj, their core-based distance E rbf The relative positions of the residues are determined by Gaussian radial basis function (RBF). lf is defined by the local frameworks created by the backbone atoms. Whether two residues touch in space is determined based on a cutoff distance between nodes. cont , the sequence relationship of the residues E sr are encoded by their relative positions.

[0151] Node feature X=[X nt , X pc , X di , X mur , X pos ] including nucleotide type X nt 、Physical and chemical properties pc And local environmental information [X di , X mur ]. Physical and chemical properties pc These include the solvent accessible surface area (SASA), which measures the extent to which nucleotides in an RNA molecule are exposed to the solvent, and the B factor. Local environmental information includes RNA surface sensing features X mur , these features are nonlinear projections of the weighted average distances from the central nucleotide to its one-hop neighbors. In addition, the dihedral angles X of the backbone atoms di describes the conformation of the RNA backbone and is represented using sine and cosine functions, X pos represents the three-dimensional coordinates of the residue. nt Diffusion is performed and denoised using the hyperbolic equivariant graph neural denoising network HEGNN proposed in this application.

[0152] Hyperbolic Position Encoding

[0153] In the above content, various encoding methods are used to represent the local environment information of the residue position. Although the local environment information is crucial for learning residue representation, the global environment information also plays an important role. In order to capture the influence of global information on residues, an improved Transformer module is used to encode the position information X of the residue. pos Encode to capture the global relationship between them and map it into the hyperbolic space. This improved Transformer module, such as Figure 3 (left side) and is defined as follows:

[0154]

[0155]

[0156] (8)

[0157] Where MHA represents the multi-head attention layer ( Figure 3 The definition of the single-head attention function is shown on the right), FFN is a feed-forward neural network.

[0158] Feature Representation in Hyperbolic Space

[0159] Hyperbolic space has higher representation power and effective distance calculation properties, making it particularly suitable for molecular embedding. First, hyperbolic space can capture complex structures in high-dimensional space in a lower dimension, thereby reducing information loss in the embedding process. This is crucial for molecules with diverse and complex structures. Second, the distance calculation properties in hyperbolic space enable it to better represent and calculate the similarity between molecules. A hyperbolic encoder is used to generate hyperbolic embeddings.

[0160] First, the Euclidean node feature x is transformed into E Mapping to the Poincare sphere .make express The origin in .

[0161]

[0162] Then, the hyperbolic point Projection to tangent space using logarithmic mapping Then, the projection result is projected onto the manifold using exponential mapping, and finally σ is used to maintain the manifold.

[0163]

[0164] Hyperbolic Equivariant Graph Neural Network Denoising Model

[0165] Biological molecules such as RNA have complex 3D structures, which are crucial to their functions. Any rotation or translation of the RNA model in three-dimensional space should not change its intrinsic properties and functions. Traditional neural networks may lose important geometric information when processing 3D data, resulting in inaccurate or unreliable predictions of RNA functions. To address this problem, this application proposes a hyperbolic equivariant graph neural network denoising model HEGNN, which utilizes the SE (3) equivariant neural layer of the equivariant graph neural network (EGNN). This method allows the HEGNN model to maintain SO (3) rotation equivariance and E (3) translation invariance when updating the representation of nodes and edges, thereby preserving geometric information and ensuring the robustness and effectiveness of the hidden representation. Specifically, in order to enable the EGNN network to update the edge representation and perform operations in hyperbolic space, a HEGNN is proposed, which includes a hyperbolic equivariant graph convolutional layer (HEGCL). HEGCL is defined as follows to update the hidden representation of a node:

[0166] To update the new edge embedding , first use the logarithmic mapping The hidden representation of the node (initially represented as x hyp ) Projection to tangent space Then, using the hyperbolic space encoding distance from formula (5) and the current edge embedding , the new edge embedding can be defined as follows:

[0167]

[0168] New hyperbolic position encoding Can be based on the current Euclidean node encoding To update:

[0169]

[0170]

[0171] Based on the above, the new node features The update can be expressed in the following form:

[0172]

[0173]

[0174] in, , , is a learnable function parameterized by a neural network, D refers to the hyperbolic space model (Poincare disk), i and j refer to the node nucleotide index, and h refers to the node hidden layer embedding (hidden embedding) (i.e., the update representation of the node embedding during the network operation process). Indicates Layer (the network is divided into multiple layers), E refers to the Euclidean space, and the corresponding and interchangeable one is the hyperbolic space (Poincare disk) referred to by D.

[0175] The embedding feature of the graph neural network node is represented by h. When the denoising process starts (that is, the initial state of the model), x D becomes x hyp (According to Formula 10), in the initial state x hyp = h D .

[0176] In the initial state x hyp = h D (At the initial t=T when denoising begins).

[0177] Training Hyperbolic Equivariant Graph Neural Network Denoising Model and DDIM Sampling

[0178] Aims to simulate the nucleotide sequence diversity of RNA while maintaining its inherent structural constraints. Train a denoising neural network to optimize the predicted probability of nucleotide type at each node , the training process is shown in Algorithm 1. The algorithm takes the RNA graph containing node features and edge features (including physicochemical properties and geometric attributes) as input. A time step t is sampled from a uniform distribution, and then the cumulative transition probability is calculated to generate noise data and embedded into the hyperbolic space. The equivariant graph neural network denoising model HEGNN predicts these noise data.

[0179] Specifically: Algorithm 1: Training

[0180] The algorithm takes an RNA graph as input, which contains node features and edge features, including physicochemical properties and geometric properties.

[0181] (1) Sample time step t from a uniform distribution;

[0182] (2) Calculate the prior distribution through the transfer matrix;

[0183] (3) Sampling a noise sequence from the distribution;

[0184] (4) Embed the noise sequence into the hyperbolic space;

[0185] (5) Predicting the probability distribution of nucleotide types using a denoising neural network;

[0186] (6) Calculate the loss function;

[0187] (7) Optimize the denoising neural network.

[0188] Since we have the prediction X described in Equation 4 nt and the posterior distribution After training, the neural network and the posterior distribution can be used to sample data. DDIM significantly accelerates the data generation process by using a non-Markov forward diffusion process, thereby reducing the number of incremental steps. By setting the noise variance to zero at each step, DDIM makes the inverse generation process completely deterministic and repeatable. The DDIM sampling process is described in Algorithm 2.

[0189] Algorithm 2: Sampling

[0190] (1) Sample a noise sequence from the prior;

[0191] (2) Iteratively denoise the noise sequence:

[0192] (3) Perform hyperbolic embedding on samples;

[0193] (4) Predicting the probability distribution of nucleotide types through a denoising neural network to obtain a predicted distribution;

[0194] (5) Calculate and update the conditional probability distribution of the state at time tk;

[0195] (6) Sampling from the conditional probability distribution of the state at time tk;

[0196] (7) After completing the iteration;

[0197] (8) When the last step is reached, the designed nucleotide sequence can be obtained by sampling the probability distribution at t = 1;

[0198] Table 1

[0199]

[0200] Table 2

[0201]

[0202] experiment

[0203] RNA reverse folding

[0204] Dataset

[0205] The model of this application was trained on the publicly available RNAsolo (Adamczyk, Antczak, and Szachniuk2022) dataset. The RNAsolo dataset is derived from a public database and contains a large number of RNA sequences and experimentally determined RNA structures. The dataset was randomly divided into an 8:1:1 ratio for training, validation, and test sets, respectively. As the sequence similarity increases, the probability of structural similarity also increases. To evaluate the generalization ability of the model, PSI-CD-HIT was used to cluster the sequences according to nucleotide similarity, and thresholds of 0.8, 0.9, and 1.0 were set.

[0206] Table 3: Sequence recovery rate (%) and F1 score (×100) at different sequence similarities (seq 0.8 / 0.9 / 1.0)

[0207]

[0208] Table 4: Comparison of recovery rate (%) and macro F1 score (×100) of sequences of different lengths at sequence similarity level 0.8

[0209]

[0210] index

[0211] The accuracy of the generated nucleotide sequences was evaluated using the recovery rate and Macro-F1 score as metrics. The recovery rate is the most commonly used metric in anti-folding, which calculates the positional consistency between the model-predicted sequence and the actual sequence. It has a higher probability of achieving correct folding of similar sequences. The F1 score is used to evaluate the accuracy and comprehensiveness of the model in predicting protein or RNA sequences by evaluating the F1 value of each letter, especially in the case of imbalanced instances of different letters.

[0212] Baseline

[0213] To evaluate the effectiveness of the proposed method, it was compared with the RNA reverse folding model and four protein tertiary structure reverse folding baselines: gRNAde (Joshi et al. 2023), PiFold (Gao, Tan, and Li 2023), StructGNN, GraphTrans, and GVP-GNN (Jing et al. 2021).

[0214] The result of reverse folding

[0215] Table 3 compares the performance of the baseline model and RI diffusion in recovering RNA sequences at different sequence similarity levels in RNAsolo. It is observed that the model of the present application achieves the best performance in terms of both recovery and F1 score. In order to conduct a more comprehensive evaluation of the model, the dataset with a sequence similarity of 0.8 was further divided into short (<50 nt), medium (50 - 100 nt) and long (>100 nt) nucleotide sequences for training. The results are shown in Table 4. It is believed that in short sequences, the lack of rich local environmental information makes it difficult for the model to distinguish different sequences; in medium-length sequences, there is sufficient and distinguishable local environmental information to capture the mapping relationship between sequence and structure; in long sequences, the complex and diverse structures that are difficult to capture and distinguish may be the reason for the low recovery rate.

[0216] Ablation studies

[0217] Ablation experiments are performed on RI diffusion, where the hyperbolic representation is replaced by the Euclidean representation and the hyperbolic position encoding is replaced by the original 3D position information. The model is trained on the RNAsolo dataset and the results are reported in Table 5.

[0218] The results show that the Transformer layer effectively captures the relationship between positions and the hyperbolic embedding significantly enhances the information representation ability of nodes in the graph.

[0219] Table 5: Ablation study results, recovery rate (%) and Macro F1 score (×100) of the model

[0220]

[0221] in conclusion

[0222] To this end, this paper proposes a geometric generation model based on RNA three-dimensional structure reverse folding - RIdiffusion. Using methods such as hyperbolic embedding and EGNN, the complex structure and geometric properties of RNA molecules are captured to recover the correct sequence from noise. The results show that the use of discrete diffusion models can significantly improve the recovery rate of RNA reverse folding and the accuracy of structural design, and outperforms the baseline method in terms of recovery rate.

[0223] Figure 4 Some results of predicting one-dimensional sequences based on a given RNA three-dimensional structure using the hyperbolic discrete diffusion model of 3D RNA reverse folding are also given, where yellow is the true (original) one-dimensional sequence of the given RNA three-dimensional structure, and green is the RNA one-dimensional sequence predicted using the hyperbolic discrete diffusion model of this application. It can be seen that the predicted one-dimensional RNA sequence is very similar to the true (original) one-dimensional sequence of the RNA. Generally, the more similar the sequences are, the more similar the structures are. A high recovery rate means that the generated sequences are more similar in structure and the design is more successful.

[0224] Accordingly, the embodiments of the present application also provide a computer-readable storage medium, in which computer executable instructions are stored, and when the computer executable instructions are executed by the processor, the various method embodiments of the present application are implemented. Computer-readable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be a computer-readable instruction, a data structure, a module of a program, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable storage media does not include temporary computer-readable media (transitory media), such as modulated data signals and carriers.

[0225] In addition, the embodiments of the present application also provide a device for predicting RNA one-dimensional sequences based on a hyperbolic discrete diffusion model, which includes a memory for storing computer executable instructions, and a processor; the processor is used to implement the steps in the above-mentioned method implementations when executing the computer executable instructions in the memory. Among them, the processor can be a central processing unit (Central Processing Unit, referred to as "CPU"), an image processor (Graphic Processing Unit, referred to as "GPU"), a digital signal processor (Digital Signal Processor, referred to as "DSP"), a microcontroller unit (Microcontroller Unit, referred to as "MCU"), a neural network processor (referred to as "NPU"), an application-specific integrated circuit (Application Specific Integrated Circuit, referred to as "ASIC"), a field programmable gate array (Field Programmable Gate Array, referred to as "FPGA") or other programmable logic devices, etc. The aforementioned memory can be a read-only memory (read-only memory, referred to as "ROM"), a random access memory (random access memory, referred to as "RAM"), a flash memory (Flash), a hard disk or a solid-state hard disk, etc. The steps of the method disclosed in each embodiment of the present invention can be directly embodied as a hardware processor to be executed, or a combination of hardware and software modules in the processor can be executed.

[0226] In addition, an embodiment of the present application further provides a computer program product, which includes computer executable instructions, and when the computer executable instructions are executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0227] The specification of this application records a large number of technical features, which are distributed in various technical solutions. If all possible combinations of technical features (i.e., technical solutions) of this application are listed, the specification will be too long. In order to avoid this problem, the various technical features disclosed in the above invention content of this application, the various technical features disclosed in the various embodiments and examples below, and the various technical features disclosed in the drawings can be freely combined with each other to form various new technical solutions (these technical solutions are all deemed to have been recorded in this specification), unless such a combination of technical features is technically infeasible. For example, in one example, feature A+B+C is disclosed, and in another example, feature A+B+D+E is disclosed. Features C and D are equivalent technical means that play the same role. Technically, only one of them can be used, and it is impossible to use them at the same time. Feature E can be combined with feature C technically. Then, the solution of A+B+C+D should not be deemed to have been recorded because it is technically infeasible, while the solution of A+B+C+E should be deemed to have been recorded.

[0228] All documents mentioned in this application are considered to be included in the disclosure of this application as a whole, so that they can be used as the basis for modification when necessary. In addition, it should be understood that after reading the above disclosure of this application, those skilled in the art can make various changes or modifications to this application, and these equivalent forms also fall within the scope of protection claimed in this application.

Claims

1. A method for constructing a hyperbolic discrete diffusion model of reverse folding of a three-dimensional RNA structure, characterized in that: The following steps are involved: (a) obtaining a three-dimensional structure of RNA and its true one-dimensional sequence, and abstractly representing the three-dimensional structure of RNA through a graph structure, wherein the graph structure includes node features and edge features of the three-dimensional structure of RNA; (b) performing a forward diffusion process on the real one-dimensional sequence of the RNA three-dimensional structure, in which noise is iteratively added to the real one-dimensional sequence of the RNA three-dimensional structure until the probability distribution of the nucleotide sequence A / U / G / C of the real one-dimensional sequence is uniformized, and outputting a noisy nucleotide sequence with a uniform probability distribution, wherein when noise is added at each time step, the transition probability at each time step is calculated; (c) randomly sampling from the noise nucleotide sequence to obtain a random noise nucleotide sequence; (d) inputting the edge features of the RNA three-dimensional structure into a multi-layer perceptron to obtain edge embedding; (e) obtaining a hyperbolic embedding of a combination of node features including the node features of the RNA three-dimensional structure and the node features of the random noise nucleotide sequence through a hyperbolic space model; (f) inputting the hyperbolic embedding of the combined node feature and the edge embedding into a hyperbolic equivariant graph neural denoising network model, training the hyperbolic equivariant graph neural network denoising model, updating the combined node feature in the hyperbolic equivariant graph neural denoising network model, and obtaining a predicted probability distribution of the intermediate nucleotide sequence based on the updated combined node feature; Based on the obtained predicted probability distribution of the intermediate nucleotide sequence and the transition probability at the corresponding time step obtained in step (b), the conditional probability distribution of the intermediate nucleotide sequence is obtained, thereby completing a denoising process; (g) adopting an intermediate one-dimensional RNA sequence from the probability distribution of the intermediate nucleotide sequence; (h) Iterating steps (e)-(g) multiple times, during the iteration, replacing the random noise nucleotide sequence in step (e) with the intermediate RNA one-dimensional sequence obtained in step (g), and after the iteration is completed, obtaining the optimal RNA one-dimensional sequence by sampling from the probability distribution of the denoised nucleotide sequence.

2. The method according to claim 1, characterized in that Step (a) also includes extracting the RNA three-dimensional skeleton structure of the RNA three-dimensional structure, wherein the RNA three-dimensional skeleton structure includes RNA skeleton atoms, and the RNA skeleton atoms include three atoms in each nucleotide, and each nucleotide is represented by the three atoms: P, C4', and N1 or N9, wherein when the nucleotide is A or G, the third skeleton atom is the atom at the N1 position, and when the nucleotide is C or U, the third skeleton atom is the atom at the N9 position, and node features and edge features are constructed through the three skeleton atoms and their relationships, the nucleotide node is represented as the C4' atom, and the edge is determined by the K proximity graph.

3. The method according to claim 2, characterized in that The node features of the RNA three-dimensional skeleton include nucleotide type, physicochemical properties, local environmental information, and the position of the nucleotide node, and the position of the nucleotide node is the three-dimensional coordinate of the nucleotide node.

4. The method according to claim 1, characterized in that In the step (f), the hyperbolic embedding of the combined node feature is updated to obtain an updated hyperbolic embedding of the combined node feature, and the updated hyperbolic embedding of the combined node feature is mapped into an updated Euclidean encoding of the combined node feature, and the predicted probability distribution of the intermediate nucleotide sequence is obtained based on the updated Euclidean encoding of the combined node feature.

5. The method according to claim 1, characterized in that In step (f), during the training of the hyperbolic equivariant graph neural network model, the edge embedding of the input Sub-step to perform update (f1): First, use a logarithmic mapping , the hidden representation of the combined node Projection to tangent space , get the combined node in Euclidean encoding of features in layer networks ; Then, the following formula is used to obtain +1 layer of updated edge embedding : Encode distances and current edge embeddings using hyperbolic space get, The updated edge embedding The definition is as follows: ， in, is a learnable function parameterized by a neural network, i and j represent the index of the combined node, where j is the neighbor node of i, Represents the network layer, h refers to the node hidden layer embedding, Indicates that the combination node i is in Euclidean coding in layer networks, It means that the combination node j is in the Euclidean encoding in layer networks, d D It represents the relative distance between the combined nodes i and j in the hyperbolic space, E represents the Euclidean space, and D represents the hyperbolic space.

6. The method according to claim 5, characterized in that The combined node feature includes the position of the nucleotide node in the node feature of the RNA three-dimensional skeleton; In step (f), during the training process of the hyperbolic equivariant graph neural network model, a sub-step (f2) is further included: updating the hyperbolic position of the nucleotide node, encoding the updated hyperbolic position of the combined node feature Based on the current Euclidean node encoding , the updated hyperbolic position encoding Updated by the following formula: , , in is a learnable function parameterized by a neural network, t represents the time step, σ is the activation function, exp c is the exponential transformation operation; Step (f3), the updated hyperbolic embedding of the combined node features is obtained by the following formula : ， , in, is a learnable function parameterized by a neural network.

7. A method for predicting one-dimensional RNA sequence based on a given three-dimensional RNA structure, characterized in that: The following steps are involved: (a) obtaining the three-dimensional structure of RNA and its true one-dimensional sequence, and extracting the RNA three-dimensional skeleton structure of the RNA three-dimensional structure, and using a k-neighbor graph G(X,E) to represent the RNA three-dimensional skeleton structure, where E is a set of edge features of the RNA three-dimensional skeleton structure, and X is a set of node features of the RNA three-dimensional skeleton structure; (b) performing a forward diffusion process on the real one-dimensional sequence of the RNA three-dimensional structure, during which noise is iteratively added to the real one-dimensional sequence of the RNA three-dimensional skeleton structure until the probability distribution of the nucleotide sequence AUGC of the real one-dimensional sequence is uniformized, and outputting a noisy nucleotide sequence with a uniform probability distribution, wherein when noise is added at each time step, the transition probability at each time step is calculated; (c) randomly sampling from the noise nucleotide sequence to obtain a random noise nucleotide sequence; (d) inputting the edge features of the RNA three-dimensional skeleton structure into a multi-layer perceptron to obtain edge embedding; (e) obtaining a hyperbolic embedding of a combination of node features including the node features of the RNA three-dimensional skeleton structure and the node features of the random noise nucleotide sequence through a hyperbolic space model; (f) during the denoising process, inputting the hyperbolic embedding of the combined node features and the edge embedding into a hyperbolic equivariant graph neural denoising network model trained according to any one of claims 1 to 6, updating the combined node features in the hyperbolic equivariant graph neural denoising network model, and obtaining a predicted probability distribution of the intermediate nucleotide sequence based on the updated combined node features; Based on the obtained predicted probability distribution of the intermediate nucleotide sequence and the transition probability at the corresponding time step obtained in step (b), the probability distribution of the intermediate nucleotide sequence is obtained, thereby completing a denoising process; (g) adopting an intermediate one-dimensional RNA sequence from the probability distribution of the intermediate nucleotide sequence; (h) Iterating steps (e)-(g), replacing the random noise nucleotide sequence in step (e) with the intermediate RNA one-dimensional sequence obtained in step (g), thereby obtaining an optimal RNA one-dimensional sequence.

8. A device for predicting one-dimensional RNA sequence based on a hyperbolic discrete diffusion model, characterized in that: include: A memory for storing computer executable instructions; as well as, A processor, coupled to the memory, is configured to implement the steps in the method of claim 7 when executing the computer executable instructions.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the steps in the method according to claim 7 are implemented.

10. A computer program product comprising computer executable instructions, characterized in that: When the computer executable instructions are executed by a processor, the steps in the method of claim 7 are implemented.

Images