Three-dimensional RNA inverse folding method for cross-graph modeling based on multi-structure comparison

Abstract

The invention discloses a cross-graph modeling three-dimensional RNA inverse folding method based on multi-structure comparison, and the method comprises the steps: carrying out the multi-structure comparison of a target RNA structure and an RNA structure database at an offline stage, obtaining a closest similar structure, carrying out the alignment processing, extracting node and edge features from the target RNA structure, and generating a training set through cross-structure coding; the training module is used for training a sequence decoder; in the online stage, through a trained sequence decoder, real-time input and cross-structure coded features are decoded, and after sequence probability distribution is obtained, a design sequence is obtained through temperature coefficient sampling. According to the method, the multi-structure comparison technology and the graph neural network are combined, and the accuracy and diversity of RNA sequence design are remarkably improved by utilizing the evolutionary conservative property of the target RNA structure and the homologous structure of the target RNA structure.

Description

Technical Field

[0001] This invention relates to a technology in the field of bioengineering, specifically a three-dimensional RNA defolding method based on cross-graph modeling (AlignIF) using multiple structure alignment. Background Technology

[0002] The three-dimensional structure of non-coding RNAs (such as aptamers, riboswitches, and ribozymes) is closely related to their biological functions. Three-dimensional RNA antifolding aims to design RNA sequences that can fold into target three-dimensional structures. However, the structural complexity and flexibility of RNA molecules make this task extremely challenging. Existing RNA design methods have limitations in capturing the dynamics and conserved structural features of RNA, often relying on real sequences or single conformations. This prevents current technologies from effectively utilizing the multiple homologous conformations that may exist in the target RNA to enhance model robustness and design accuracy, and also limits their ability to design target RNAs lacking protein ligands. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention proposes a three-dimensional RNA defolding method based on cross-graph modeling using multi-structure alignment. By combining multi-structure alignment technology and graph neural networks, and utilizing the evolutionary conservation of the target RNA structure and its homologous structures, a graph representation of the target RNA integrating multi-conformation conserved features is generated. This fully preserves structural commonalities and spatial constraints, providing a high-fidelity structural prior for subsequent RNA sequence defolding tasks, and significantly improving the accuracy and robustness of defolding.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a method for three-dimensional RNA defolding based on cross-graph modeling using multi-structure alignment. First, the three-dimensional structure of the target RNA is compared with an RNA structure database to identify candidate structures. Then, the target structure is structurally aligned with each candidate structure, and the target RNA and its multiple similar structures are modeled as geometric graphs, from which node and edge features are extracted. Next, a three-layer cross-structure encoder is used to jointly encode the geometric graphs. Through information interaction and feature aggregation, a semantic representation of the target structure is obtained. A sequence decoder then generates a graph representation of the target RNA incorporating conserved features across multiple conformations. In the online phase, the trained three-layer cross-structure encoder generates node-level feature representations rich in multi-conformation information based on the input three-dimensional structure of the target RNA. The sequence decoder then decodes the conditional probability distribution of each nucleotide site (A, G, C, U) position by position in an autoregressive manner. Based on this probability distribution, combined with a sampling strategy with an adjustable temperature coefficient, diverse and structurally compatible RNA sequence design schemes can be generated.

[0006] The RNA structure database mentioned includes, but is not limited to, the Protein Database (PDB), which contains a large number of experimentally determined three-dimensional RNA structures, covering a variety of functional categories and folding types, providing rich and high-quality structural resources for multi-structure comparison.

[0007] The aforementioned multi-structure alignment and structural alignment are achieved using, but not limited to, the US-align tool, which can accurately align RNA structures and output alignment results.

[0008] The node features include: dihedral angle features, orientation features, and relative rotation features, wherein: dihedral angle features describe the local conformation of nucleotides, including seven natural dihedral angles (α, β, γ, δ, ε, ζ, ...). The x-axis and the two internal angles of the sugar ring (C2'-C1'-O4'-C4', O4'-C1'-C2'-O2') are both encoded with sine and cosine. The directional features describe the spatial orientation between key atoms (such as P→C4', base N1 / N9→C4'), which are represented in a local coordinate system constructed by O4', C4', and C3': the x-axis is the bisector of the angle between O4'→C4' and C3'→C4', the y-axis is (O4'→C4')×(C3'→O4'), and the z-axis is x×y, ensuring rotation invariance. The relative rotation features describe the spatial orientation relationship between the current nucleotide and its neighboring nucleotides, which are quantized by the rotation matrix between the local coordinate systems and converted into quaternion encoding to achieve a compact, singularity-free representation.

[0009] The edge features include distance features and relative rotation features, wherein: the distance features include Euclidean distances between C4'-C4', P-C4', N1 / N9-C4', and N1 / N9-N1 / N9, and are encoded using radial basis functions; the relative rotation features represent the spatial rotation between the local coordinate systems of the two nucleotides using quaternions, characterizing their relative orientation.

[0010] The cross-structure encoder and sequence decoder, during the offline training phase, utilize known RNA structure-sequence pairing data to construct a dataset based on structural similarity. After joint training in an end-to-end manner, the cross-structure encoder and sequence decoder run cyclically to finally output a complete RNA sequence, achieving structure-aware unbiased autoregressive generation.

[0011] The cross-structure encoder includes: an intra-graph feature aggregation module, a cross-graph attention calculation module, a context feature fusion module, and a multi-source feature integration module. Specifically: the intra-graph feature aggregation module updates the feature representations of nodes and edges for each independent input geometric graph (including the target structure graph and similar structure graphs obtained through structure comparison), utilizing the features of the current node, the features of connected edges, and information from neighboring nodes, through a learnable message passing function and aggregation function; the cross-graph attention calculation module fuses evolutionarily conserved information from topological alignment positions in multiple similar structure graphs while maintaining structural arrangement invariance; the context feature fusion module is based on attention weights. For nodes in similar structure graphs value vector (Obtained from its features through linear transformation) Weighted aggregation is performed to generate the target node. Cross-graph context vector The multi-source feature integration module will integrate the context vector. With the target node The features obtained after message passing within the graph are fused together and used as input features for the next layer.

[0012] The sequence decoder includes an encoding module, an indexing module, a distribution prediction module, and a base generation module. The encoding module performs graph neural network (GNN) encoding based on the node features of the target RNA structure map and the generated sequence information to obtain the structure-sequence fusion features of the 128-dimensional hidden space. The indexing module performs random position selection based on the position information of the currently ungenerated sequence to obtain the index result of the site to be predicted, t. The distribution prediction module performs conditional probability calculation based on the GNN encoding features and the structural context information of site t to obtain the nucleotide distribution prediction result for that site. The base generation module performs nucleotide sampling based on the predicted probability distribution information to generate the final bases of site t and update the sequence state.

[0013] The joint training process introduces a randomized decoding order strategy to enhance the model's adaptability to different sequence generation paths and improve its robustness and generalization performance; at the same time, it supports using known subsequences as prior information input to achieve conditional sequence design.

[0014] The joint training employs label-smoothed cross-entropy loss with a smoothing parameter set to 0.1 to measure the difference between the model's predicted sequence and the true label sequence, thereby improving the model's generalization ability, effectively mitigating overfitting, and enhancing the quality of the model's output probability distribution. The Adam optimizer is used for parameter updates with a learning rate of 1e-4 and a training batch size of 32. When training the sequence decoder, the nucleotides from the previous time step in the true target sequence are used as the input for the current time step, rather than using the model's own predictions. This strategy helps accelerate model convergence and enhances the stability of the training process.

[0015] This invention relates to a system for implementing the above-mentioned method, comprising: a multi-structure alignment module, a feature extraction module, a cross-graph modeling module, a sequence decoding module, a training module, and a data storage module, wherein: the multi-structure alignment module, based on the three-dimensional coordinate data of the target RNA structure, calls the US-align tool to compare with an RNA structure database, and filters out multiple similar structures and their alignment information with the target structure; the feature extraction module, based on the alignment information, extracts node features and edge features from the target RNA structure, constructs the target geometry, and simultaneously constructs a geometry for each similar structure obtained from the alignment; the cross-graph modeling module, based on the target geometry and multiple similar structure geometry, uses... Through iterative processing involving intra-graph message passing and cross-graph self-attention integration, the system outputs a target graph representation incorporating multi-conformation information. The sequence decoding module predicts the nucleotide probability distribution step-by-step in random order based on the target graph representation and generates the final designed sequence using temperature sampling. The training module manages the offline training process, loads the training dataset (target graph and its real sequences), processes the input graph using random masking techniques, calls the cross-graph modeling module to process the graph data, calls the sequence decoder for prediction, calculates the cross-entropy loss, and updates the parameters of the cross-graph modeling module and the sequence decoder through backpropagation. The data storage and management module stores the RNA structure database, training dataset, model parameters, and design results.

[0016] Technical effect

[0017] This invention constructs and collaboratively learns graph structures corresponding to multiple similar RNA structures. Each graph uses nucleotides as nodes and spatial interactions as edges to characterize the topological relationships of RNA under different potential conformations, thereby explicitly modeling its structural flexibility. At the graph construction level, this invention designs a high-dimensional geometric feature encoding mechanism. Node features include the sine and cosine values ​​of the dihedral angles of nucleotides, describing local geometric configurations; directional features are calculated using the direction vectors between phosphate groups, sugar groups, and base groups; relative rotation features use quaternions to represent the spatial relationships between adjacent nucleotides. Edge features fuse the distance and directional relationships between nucleotides, enhancing the ability to perceive long-range interactions. Furthermore, this invention proposes a multi-graph feature fusion framework based on a self-attention mechanism to jointly update the node and edge features of multiple conformational graphs. This mechanism not only captures long-range dependencies within a single graph, but more importantly, establishes cross-graph attention connections between different conformational graphs, realizing information interaction and knowledge sharing between potential conformations, thereby improving the model's adaptability to RNA structural diversity. To improve the model's generalization ability and integrate prior knowledge of known important subsequences, this invention introduces a random sequence decoding module that supports sequence decoding with arbitrary nucleotide arrangements, thereby enhancing the model's robustness and perception of sequence diversity.

[0018] Compared with existing technologies, this invention significantly improves the accuracy of RNA sequence design. On the baseline test set, this method achieves a 60% sequence recovery rate, a 33.3% improvement over the current best-in-class technology; on the independent test set, the sequence recovery rate reaches 48%, a 16.7% improvement over existing methods (see...). Figure 3 The above results fully demonstrate the superior ability of this method to capture complex RNA structural patterns, showcasing stronger generalization performance and practical application potential. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention;

[0020] Figure 2 This is a schematic diagram of the invention;

[0021] Figure 3 This is a schematic diagram illustrating the effect of an example. Detailed Implementation

[0022] like Figure 1 As shown, this embodiment relates to a three-dimensional RNA defolding method based on multi-structure alignment and cross-graph modeling, including:

[0023] Step 1, Structure Alignment: Given a target RNA structure, the US-align tool is used to align it with all structures in the RNA structure database. To ensure similarity in global folding conformations, matching structures with a TM-score greater than 0.45 are selected, thus obtaining all RNA structures that are topologically similar. US-align also provides detailed alignment information for subsequent cross-graph coding.

[0024] The target RNA structure refers to the RNA molecular structure to be analyzed provided by the user.

[0025] Step 2, Graph Representation Construction and Feature Extraction: All RNA structures are represented as geometric graphs, with each node representing a nucleotide, located by the three-dimensional coordinates of its C4' atom, and any two nodes within a spatial distance of 20 Å are connected by edges.

[0026] The node features include dihedral angle features, orientation features and relative rotation features in the local coordinate system, and a spatial rotation matrix based on quaternions.

[0027] The dihedral features described are used to describe the local geometric conformation of nucleotides, including seven naturally occurring nucleotide dihedrals (α, β, γ, δ, ε, ζ, χ) and two internal dihedrals of the sugar ring proposed in this invention (C2'-C1'-O4'-C4' and O4'-C1'-C2'-O2') to more accurately capture the dynamic changes in sugar ring conformation. All dihedrals are encoded using sine and cosine functions.

[0028] The directional features are used to characterize the spatial orientation relationships between key atoms, specifically including: the direction from the P atom in the phosphate group to the C4' atom in the sugar ring, and the direction from the N1 (pyrimidine) or N9 (purine) atom in the base to the C4' atom. To ensure rotational invariance, these direction vectors are represented in a local coordinate system defined by each nucleotide. The local coordinate system proposed in this invention is constructed from three atoms O4', C4', and C3' on the sugar ring, which is more stable and adaptable to the flexibility of the RNA chain compared to traditional systems based on continuous phosphorus atoms. The x-axis is defined as the direction of the angle bisector of the angle between the two vectors O4'→C4' and C3'→C4', the y-axis is determined by the cross product of the vectors O4'→C4' and C3'→O4', and the z-axis is obtained by the cross product of the x-axis and y-axis, thus forming an orthogonal coordinate system.

[0029] The relative rotation feature is used to describe the spatial orientation relationship between the current nucleotide and its neighboring nucleotides (along the 3' and 5' directions). This relationship is quantified by a relative rotation matrix between the local coordinate systems of the neighboring nucleotides, and the rotation matrix is ​​encoded in quaternion form to achieve a compact and singularity-free rotation representation.

[0030] The edge features include distance features and relative rotation features.

[0031] The features described cover the Euclidean distances between multiple key atom pairs, specifically the distances between four key atom pairs: C4'-C4', P-C4', N1 / N9-C4', and N1 / N9-N1 / N9, and these inter-atom distances are encoded using radial basis functions (RBF).

[0032] The relative rotation feature in the edge features represents the spatial rotation relationship between the local coordinate systems of the two nucleotides. The corresponding rotation matrix is ​​calculated and encoded into a quaternion form to characterize the relative orientation between the nucleotides.

[0033] Step 3: Construct a three-layer cross-structure encoder. Each layer of the cross-structure encoder includes: an intra-graph feature aggregation module, a cross-graph attention calculation module, a context feature fusion module, and a multi-source feature integration module. Specifically: the intra-graph feature aggregation module updates the feature representations of nodes and edges for each independent input geometric graph (including the target structure graph and similar structure graphs obtained through structure comparison), utilizing the features of the current node, the features of connected edges, and information from neighboring nodes, through learnable message passing functions and aggregation functions; the cross-graph attention calculation module fuses evolutionarily conserved information from topological alignment positions in multiple similar structure graphs while maintaining the permutation invariance of the structure; the context feature fusion module is based on attention weights... For nodes in similar structure graphs value vector (Obtained from its features through linear transformation) Weighted aggregation is performed to generate the target node. Cross-graph context vector The multi-source feature integration module will integrate the context vector. With the target node The features obtained after message passing within the graph are fused together and used as input features for the next layer.

[0034] In this embodiment, each layer of node vectors uses a 128-dimensional hidden feature space to balance expressive power and computational efficiency.

[0035] The fusion, specifically, refers to: for each node in the target graph... Calculate its corresponding topological position nodes in all similar structural graphs. attention weights ,in: For the target node The query vector generated by linear transformation of the features. For corresponding nodes in similar structures The key vector obtained by linear transformation of the features. Represents the vector dimension.

[0036] The multi-source feature integration module fuses features resulting from intra-graph message passing and cross-graph feature integration, enabling node features to propagate fully not only within their local neighborhoods but also to complete multi-level information interaction between cross-graph alignment positions. This dual propagation mechanism significantly enhances the model's ability to characterize the geometric properties of the target structure itself and its ability to jointly model evolutionary conservatism patterns among multiple structures. After three layers of progressive feature extraction, high-order feature representations of each node in the target graph are finally generated. These representations deeply integrate topological consistency information from multiple alignment structures. The fused node features, as the overall semantic representation of the target structure, will be used in subsequent sequence decoding tasks, providing strong semantic support for structure-aware sequence generation.

[0037] Step 4: Construct a sequence decoder, which specifically includes an encoding module, an indexing module, a distribution prediction module, and a base generation module. Specifically: the encoding module performs graph neural network (GNN) encoding based on the node features of the target RNA structure map and the generated sequence information to obtain the structure-sequence fusion features of the 128-dimensional hidden space; the indexing module performs random position selection based on the position information of the currently ungenerated sequence to obtain the index result of the site to be predicted, t; the distribution prediction module performs conditional probability calculation based on the GNN encoding features and the structural context information of site t to obtain the nucleotide distribution prediction result of the site; and the base generation module performs nucleotide sampling based on the predicted probability distribution information to generate the final bases of site t and update the sequence state.

[0038] The sequence decoder iteratively performs random sequential sampling, autoregressive decoding, and nucleotide generation based on the node feature representation of the target RNA structure map output by the cross-structure encoder and the currently generated RNA sequence (initially empty), ultimately outputting a complete RNA sequence. Specifically, this includes:

[0039] a) In each decoding step, instead of using a traditional fixed order (e.g., from the 5' to the 3' end), a target position t is randomly selected from all positions where nucleotides have not yet been generated for prediction. This strategy breaks the directional bias in sequence generation, enhances the model's ability to model global structural dependencies, and improves the diversity and robustness of generated sequences.

[0040] b) Autoregressive Decoding: For a selected position t, the current graph structure representation is combined with the nucleotide type features of each position in the generated sequence and input into the GNN. Through multi-layer message passing and feature aggregation operations, the model captures the generated context. With structural features The complex dependencies between them. Ultimately, the conditional probability distributions of the four nucleotides (A, G, C, U) at position t are predicted by the output layer (a linear transformation layer followed by a softmax function). .

[0041] c) Based on the predicted probability distribution, a sampling strategy with a temperature parameter is used to generate the nucleotide corresponding to position t and add it to the generated sequence.

[0042] d) Repeat step ac until nucleotides at all positions are generated, and finally output a complete RNA sequence.

[0043] Step 5: Training the cross-structure encoder and sequence decoder: Training was performed using a dataset of 7,932 RNA structures from the PDB database. The sequence decoder was trained in a self-supervised manner, employing cross-entropy loss (label smoothing value of 0.1) and a teacher-forced strategy, achieving stable and rapid convergence by inputting the previous true label. All structures had a resolution no higher than 4 Å and sequence lengths between 16 and 512 nucleotides. The samples were clustered based on a TM-score threshold of 0.45, resulting in 499 clusters. These clusters were then divided into a training set (399 clusters, 7,669 structures), a validation set (50 clusters, 129 structures), and a test set (50 clusters, 158 structures) in an 8:1:1 ratio. Model training used a labeled smoothed cross-entropy loss function (smoothing coefficient of 0.1), and early stopping was determined by monitoring the loss changes on the validation set to effectively prevent overfitting and improve model generalization ability.

[0044] Through specific practical experiments, this invention significantly improves the sequence recovery rate. Figure 3 The perplexity and Macro-F1 score are shown in Table 1.

[0045] Table 1

[0046] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A three-dimensional RNA defolding method based on cross-graph modeling using multiple structure alignments, characterized in that, First, the three-dimensional structure of the target RNA is compared with the RNA structure database to select candidate structures. Then, the target structure is aligned with each candidate structure. The target RNA and the multiple similar structures obtained from the comparison are modeled as geometric graphs, and node and edge features are extracted from them. The geometric graph is then jointly encoded using a three-layer cross-structure encoder. The semantic representation of the target structure is obtained through information interaction and feature aggregation. The target RNA graph representation, which incorporates conserved features of multiple conformations, is then generated by a sequence decoder. In the online stage, the trained three-layer cross-structure encoder generates node-level feature representations rich in multiple conformation information based on the three-dimensional structure of the input target RNA. The conditional probability distribution of each nucleotide site (A, G, C, U) is then decoded position by position by the sequence decoder in an autoregressive manner.

2. The three-dimensional RNA defolding method based on cross-graph modeling with multi-structure alignment as described in claim 1, characterized in that, The node features include: dihedral angle features, orientation features, and relative rotation features, wherein: dihedral angle features describe the local conformation of nucleotides, including seven natural dihedral angles (α, β, γ, δ, ε, ζ, ...). The x-axis and the two internal angles of the sugar ring (C2'-C1'-O4'-C4', O4'-C1'-C2'-O2') are both encoded with sine and cosine. The directional features describe the spatial orientation between key atoms (such as P→C4', base N1 / N9→C4'), which are represented in a local coordinate system constructed by O4', C4', and C3': the x-axis is the bisector of the angle between O4'→C4' and C3'→C4', the y-axis is (O4'→C4')×(C3'→O4'), and the z-axis is x×y, ensuring rotation invariance. The relative rotation features describe the spatial orientation relationship between the current nucleotide and its neighboring nucleotides, which are quantized by the rotation matrix between the local coordinate systems and converted into quaternion encoding to achieve a compact, singularity-free representation.

3. The three-dimensional RNA defolding method based on cross-graph modeling with multi-structure alignment as described in claim 1, characterized in that, The edge features include distance features and relative rotation features, wherein: the distance features include Euclidean distances between C4'-C4', P-C4', N1 / N9-C4', and N1 / N9-N1 / N9, and are encoded using radial basis functions; the relative rotation features represent the spatial rotation between the local coordinate systems of the two nucleotides using quaternions, characterizing their relative orientation.

4. The three-dimensional RNA defolding method based on cross-graph modeling with multiple structure alignment as described in claim 1, characterized in that, The cross-structure encoder and sequence decoder, during the offline training phase, utilize known RNA structure-sequence pairing data to construct a dataset based on structural similarity. After joint training in an end-to-end manner, the cross-structure encoder and sequence decoder run cyclically to finally output a complete RNA sequence, achieving structure-aware unbiased autoregressive generation.

5. The three-dimensional RNA defolding method based on cross-graph modeling with multiple structure alignment as described in claim 1 or 4, characterized in that, The cross-structure encoder includes: an intra-graph feature aggregation module, a cross-graph attention calculation module, a contextual feature fusion module, and a multi-source feature integration module. Specifically: the intra-graph feature aggregation module updates the feature representations of nodes and edges for each independent input geometric graph using the features of the current node, the features of connected edges, and information from neighboring nodes, through a learnable message passing function and aggregation function; the cross-graph attention calculation module fuses evolutionarily conserved information from topological alignment positions in multiple similar structural graphs while maintaining structural arrangement invariance; the contextual feature fusion module is based on attention weights. For nodes in similar structure graphs value vector Perform weighted aggregation to generate the target node. Cross-graph context vector The multi-source feature integration module will integrate the context vector. With the target node The features obtained after message passing within the graph are fused together and used as input features for the next layer.

6. The three-dimensional RNA defolding method based on cross-graph modeling with multiple structure alignment as described in claim 1 or 4, characterized in that, The sequence decoder includes an encoding module, an indexing module, a distribution prediction module, and a base generation module. The encoding module performs graph neural network (GNN) encoding based on the node features of the target RNA structure map and the generated sequence information to obtain the structure-sequence fusion features of the 128-dimensional hidden space. The indexing module performs random position selection based on the position information of the currently ungenerated sequence to obtain the index result of the site to be predicted, t. The distribution prediction module performs conditional probability calculation based on the GNN encoding features and the structural context information of site t to obtain the nucleotide distribution prediction result for that site. The base generation module performs nucleotide sampling based on the predicted probability distribution information to generate the final bases of site t and update the sequence state.

7. The three-dimensional RNA defolding method based on cross-graph modeling with multi-structure alignment according to claim 4, characterized in that, The joint training process introduces a randomized decoding order strategy to enhance the model's adaptability to different sequence generation paths, improving its robustness and generalization performance. It also supports using known subsequences as prior information input to achieve conditional sequence design. Specifically, it employs label-smoothed cross-entropy loss with a smoothing parameter set to 0.1 to measure the difference between the model's predicted sequence and the true label sequence, thereby improving the model's generalization ability, effectively mitigating overfitting, and enhancing the quality of the model's output probability distribution. The Adam optimizer is used for parameter updates with a learning rate of 1e-4 and a training batch size of 32. When training the sequence decoder, the nucleotides from the previous time step in the real target sequence are used as the input for the current time step, rather than using the model's own predictions.

8. A three-dimensional RNA defolding system for cross-graph modeling implementing the method of any one of claims 1-7, characterized in that, include: The system comprises a multi-structure alignment module, a feature extraction module, a cross-graph modeling module, a sequence decoding module, a training module, and a data storage module. Specifically: the multi-structure alignment module uses the US-align tool to align the target RNA structure with an RNA structure database based on its 3D coordinate data, selecting multiple similar structures and their alignment information with the target structure; the feature extraction module extracts node and edge features from the target RNA structure based on the alignment information, constructing a target geometry while simultaneously building a geometry for each similar structure obtained from the alignment; the cross-graph modeling module outputs a target graph representation incorporating multi-conformation information through iterative processing of the target geometry and multiple similar structure geometry through intra-graph message passing and cross-graph self-attention integration; the sequence decoding module predicts the nucleotide probability distribution step-by-step in random order based on the target graph representation and applies temperature sampling to generate the final designed sequence; the training module manages the offline training process, loads the training dataset, applies random masking techniques to process the input graph, calls the cross-graph modeling module to process the graph data, calls the sequence decoder for prediction, calculates cross-entropy loss, and backpropagates to update the parameters of the cross-graph modeling module and the sequence decoder; and the data storage and management module stores the RNA structure database, training dataset, model parameters, and design results.

Images