RNA residue contact prediction method and system based on sandglass type residual coding and decoding network

Abstract

The application belongs to the technical field of bioinformatics and computational biology, and discloses an RNA residue contact prediction method and system based on an hourglass residual coding and decoding network. The application first completes vectorization representation based on sequence composition features, and then divides an RNA sequence database into multiple low-redundancy independent sequence clusters through a clustering algorithm. The optimal matching covariance model of a target RNA sequence in an Rfam database is taken as a guide to independently complete homologous sequence searching in each sequence cluster, and multiple sets of independent sub-multisequence alignments are constructed. Evolutionary coupling feature matrices of each sub-alignment are extracted through direct coupling analysis, spliced into a multi-channel feature tensor, and then input into a designed hourglass residual coding and decoding network to complete feature fusion and contact prediction. Finally, a contact probability matrix between RNA residues is output through symmetrization processing. The application can significantly improve the accuracy of RNA residue contact prediction and provide reliable spatial constraints for de novo modeling of RNA three-dimensional structures.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and computational biology, specifically relating to a method and system for predicting RNA residue contacts based on an hourglass-shaped residual encoding and decoding network. Background Technology

[0002] The biological functions of RNA molecules depend on their three-dimensional spatial structure. However, due to the high cost and complexity of experimental techniques such as X-ray crystallography, nuclear magnetic resonance, and cryo-electron microscopy, the number of resolved RNA three-dimensional structures is far less than the scale of their sequence data, severely restricting in-depth research on RNA functional mechanisms. Therefore, developing efficient and accurate computational prediction methods based on sequence information has become an important research direction in this field.

[0003] RNA residue contact prediction aims to infer the spatial proximity between nucleotide pairs from sequence information. Accurate contact prediction can provide spatial constraints for three-dimensional structure modeling. Existing RNA residue contact prediction methods have the following shortcomings:

[0004] First, homology searches are inefficient and the quality of multiple sequence alignments is unstable. Existing methods typically perform global homology searches directly in large-scale RNA sequence databases without preprocessing, resulting in high computational complexity. Furthermore, the search results are mixed with a large number of phylogenetic noise sequences, leading to inconsistent quality of the constructed multiple sequence alignments and the extraction of co-evolutionary signals containing a large amount of indirect noise.

[0005] Second, homology searches lack reliable prior guidance, and single alignments mask differentiated evolutionary patterns. Existing methods either automatically construct covariance models or extract evolutionary coupling features after constructing single multiple sequence alignments on the entire database. This process mixes homologous sequences from different phylogenetic origins, averaging the co-evolutionary signals of different evolutionary branches and weakening the predictive ability for spatial contact of RNA structures.

[0006] Third, existing deep learning methods employ a single-channel input strategy to process evolutionarily coupled feature matrices, lacking an effective fusion mechanism for multi-channel differential features. Furthermore, convolutional networks may introduce directional prediction biases during unidirectional inference, failing to satisfy the undirected symmetry constraint of spatial interactions between RNA molecule residues.

[0007] Therefore, existing technologies have significant shortcomings in terms of homology search efficiency, multiple sequence alignment quality, utilization of differential evolutionary features, and symmetry constraints on output results. There is an urgent need for an RNA residue contact prediction method that can simultaneously improve computational efficiency and prediction accuracy. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, such as low homology search efficiency, unstable quality of multiple sequence alignment, lack of fusion mechanism for multi-channel differential evolutionary features, and directional bias in network inference, this invention provides an RNA residue contact prediction method and system based on an hourglass-shaped residual encoding and decoding network.

[0009] The technical solution adopted in this invention is: an RNA residue contact prediction method based on an hourglass-shaped residual encoder-decoder network, comprising the following steps:

[0010] S1: Based on the sequence composition features of RNA sequences, the nucleotide sequences in the RNA sequence database are vectorized and then the RNA sequence database is divided into multiple independent sequence clusters with low redundancy through a clustering algorithm.

[0011] S2: Obtain the target RNA sequence, calculate and determine the optimal covariance model of the target RNA sequence in the Rfam known RNA family database, and use the covariance model as a guide to perform homologous sequence search in each independent sequence cluster obtained in step S1, and construct the sub-multiple sequence alignment corresponding to each independent sequence cluster.

[0012] S3: Perform direct coupling analysis on each sub-sequence alignment and extract the corresponding evolutionary coupling feature matrix;

[0013] S4: Concatenate multiple evolutionary coupling feature matrices into a multi-channel input feature tensor, input it into an hourglass-shaped residual encoder-decoder network for feature fusion and contact prediction, and output the RNA residue contact probability matrix of the target RNA sequence after symmetry processing.

[0014] Furthermore, the vectorized representation based on sequence composition features described in step S1 constructs a fixed-length feature vector by statistically analyzing the occurrence frequency of each triplet fragment in the nucleotide sequence; the clustering algorithm is based on the L2 norm distance metric between feature vectors, grouping sequences with similar composition features into the same sequence cluster.

[0015] Furthermore, in step S2, the covariance model is determined and the sub-multiple sequence alignment is constructed, specifically including:

[0016] The matching scores of the target RNA sequence with each covariance model in the Rfam family database are calculated, and the covariance model with the highest score is selected as the guiding model for homology search. Based on the guiding model, homology sequence search is performed independently in each sequence cluster, and an E-value threshold is set to filter homology sequences. For the retrieved homology sequences, sequence alignment guided by the covariance model is performed with the target RNA sequence as a reference. All alignment columns corresponding to the gap positions in the target sequence are pruned so that the column dimension of each sub-multiple sequence alignment is consistent with the length of the target RNA sequence, and sub-multiple sequence alignments corresponding to each sequence cluster are generated respectively.

[0017] Furthermore, the direct coupling analysis method described in step S3 is an analysis method for extracting direct evolutionary coupling relationships between residues in multiple sequence alignments, including a pseudo-likelihood maximization direct coupling analysis method; by independently processing each of the sub-multiple sequence alignments, corresponding... The evolutionary coupling feature matrix; the multi-channel input feature tensor mentioned in step S4, which consists of multiple... Constructed by splicing together evolutionary coupling feature matrices, with dimension 1. Where N is the number of sequence clusters, The length of the target RNA sequence is nucleotides.

[0018] Furthermore, the hourglass-shaped residual codec network architecture described in step S4 specifically includes:

[0019] The encoder module consists of cascaded residual convolutional blocks, which perform multi-level feature abstraction on the multi-channel feature tensor. The number of feature channels gradually increases along the network depth direction, and outputs a high-dimensional abstract feature map.

[0020] Decoder module: Composed of cascaded residual convolutional blocks symmetrical to the encoder module, it performs feature reconstruction and spatial resolution restoration on the high-dimensional abstract feature map. The number of feature channels is progressively compressed along the network depth direction to realize feature reconstruction and prediction of contact probability between residues.

[0021] Single-channel output module: The multi-channel feature map output by the decoder module is reduced to a single channel through convolution, and the predicted value is mapped to the [0,1] probability interval by the Sigmoid activation function, and the contact probability matrix between RNA residues of the target RNA sequence is output.

[0022] Furthermore, the encoder module contains two cascaded residual convolutional blocks, each consisting of a main branch and a residual branch: the main branch uses a double-layer 3×3 convolutional kernel with an equal-size edge padding strategy to expand the receptive field while maintaining the feature map spatial resolution; the residual branch achieves linear transformation and dimension matching across channel dimensions through 1×1 convolution; the outputs of the two branches are added element-wise and then activated by the LeakyReLU activation function to form a complete residual mapping, which is then introduced into a two-dimensional random deactivation layer after the encoder module output;

[0023] The weight initialization strategy of the hourglass-shaped residual codec network is as follows: all two-dimensional convolutional layers are initialized with a Kaiming normal distribution to maintain the stability of the signal variance during forward propagation; the final single-channel output layer is initialized with a Xavier uniform distribution.

[0024] Furthermore, an absolute error loss function is constructed to train the hourglass-shaped residual encoder-decoder network. The mathematical expression of the absolute error loss function is:

[0025]

[0026] Where N is the total number of valid elements in the matrix; The predicted contact probability of residue pair (i, j) output by the network; ∈{0,1} represents a true binary contact label; The asymmetric penalty weight is defined as follows:

[0027]

[0028] in, The penalty weight hyperparameter for positive samples greater than 1.

[0029] Furthermore, after outputting the RNA residue contact probability matrix in step S4, a symmetry processing step is also included, specifically:

[0030] For the predicted probability values ​​of the two symmetrical positions (i, j) and (j, i) representing the same nucleotide pair (i, j) in the contact probability matrix, the larger of the two values ​​is taken as the final spatial contact probability of the nucleotide pair, and the corresponding two symmetrical positions in the matrix are updated accordingly to generate a symmetrical final contact probability matrix.

[0031] The symmetry processing eliminates the directional prediction bias generated during unidirectional convolution inference by integrating the prediction results of symmetrical positions in the matrix, so that the output matrix conforms to the undirected symmetry constraint of spatial interaction between RNA residues.

[0032] Furthermore, after outputting the final contact probability matrix, the continuous probability values ​​are binarized into contact determination results. Based on the principle of maximizing precision, recall, F1 score, or Hughes correlation coefficient, the contact determination threshold is set to... That is, when the predicted probability is not lower than If the nucleotide pair is in spatial contact, it is determined that there is contact; otherwise, it is determined that there is no contact.

[0033] This invention also provides an RNA residue contact prediction system based on an hourglass-shaped residual encoder-decoder network, comprising:

[0034] The sequence clustering module is used to vectorize the nucleotide sequences in the RNA sequence database based on the sequence composition features of RNA sequences, and then divide the RNA sequence database into multiple independent sequence clusters with low redundancy through a clustering algorithm.

[0035] The homology search module is used to obtain the target RNA sequence, calculate and determine the optimal covariance model of the target RNA sequence in the Rfam known RNA family database, and use the covariance model as a guide to perform homology sequence search in each independent sequence cluster obtained by the sequence clustering module, and construct the sub-multiple sequence alignment corresponding to each independent sequence cluster.

[0036] The coupling feature extraction module is used to perform direct coupling analysis on each sub-sequence alignment and extract the corresponding evolutionary coupling feature matrix.

[0037] The multi-channel fusion and prediction module is used to concatenate multiple evolutionary coupling feature matrices into a multi-channel input feature tensor, which is then input into an hourglass-shaped residual encoder-decoder network for feature fusion and contact prediction. After symmetry processing, the module outputs the RNA residue contact probability matrix of the target RNA sequence.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] First, it improves the efficiency of homology search and the quality of multi-sequence alignment. This invention transforms the global search into a separate homology search within each independent sequence cluster by clustering the homology search database based on sequence composition features. This effectively filters out heterogeneous sequences with significant compositional differences, reduces computational complexity, and ensures the construction quality of each sub-multi-sequence alignment from the source.

[0040] Second, it preserves multi-source differentiated evolutionary signals. This invention adopts a strategy of constructing sub-sequence alignments independently in clusters and extracting coupling features separately, which avoids the mutual masking of features between different homologous subsets, so that the multi-channel feature tensor retains the independent evolutionary coupling signals of each subset and improves the information representation ability of evolutionary features.

[0041] Third, it enhances the ability to capture multi-level features. The hourglass-shaped residual encoder-decoder network designed in this invention achieves multi-level feature abstraction through progressive channel amplification on the encoder side, and achieves spatial resolution restoration through progressive channel compression on the decoder side, effectively capturing multi-level structural constraint information of RNA residue contacts.

[0042] Fourth, it satisfies the undirected symmetry constraint. The symmetry processing mechanism introduced in this invention eliminates the directional bias in the unidirectional convolution inference process by integrating the prediction results of symmetrical positions in the matrix, so that the prediction results conform to the undirected symmetry constraint of the interaction between RNA molecule residues, providing spatial constraint information for downstream three-dimensional structure modeling. Attached Figure Description

[0043] Figure 1 This is a diagram of the hourglass-shaped residual encoding / decoding network framework in an embodiment of the present invention;

[0044] Figure 2 This is a bar chart showing the hierarchical prediction performance evaluation results based on sequence interval distance on an independent test set in an embodiment of the present invention.

[0045] Figure 3 This is a radar chart comparing the prediction performance of the MSA strategy and the MSAs strategy on an independent test set in this embodiment of the invention, covering the entire range ( >0), medium-to-long-range contact ( >4) and remote topology contact ( >24) Subgraphs of the distance intervals between the three sequences.

[0046] Figure 4 This is a schematic diagram illustrating the contact prediction effect of RNA structure 1Y26 chain A in an embodiment of the present invention; wherein... Figure 4 In the diagram, A represents a two-level topology. Figure 4 In the diagram, B represents the corresponding three-level structure visualization. Figure 4 In the diagram, C represents a comparison between the predicted probability matrix and the actual contact labels. The upper triangle represents the predicted probability distribution, and the lower triangle represents the actual contact labels. Detailed Implementation

[0047] To facilitate understanding and implementation of the technical solutions of this invention by those skilled in the art, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the scope of protection of this invention.

[0048] This invention provides a method for predicting RNA residue contacts based on an hourglass-shaped residual encoder-decoder network, the overall implementation framework of which is as follows: Figure 1 As shown, the process mainly includes four stages: dataset construction and quality control, homologous sequence clustering and feature extraction (steps S1 to S3), and multi-channel network inference and symmetry processing (step S4 and subsequent processing). The specific implementation process is as follows.

[0049] This embodiment constructs a benchmark dataset based on publicly available RNA structure databases. Through quality control steps including resolution filtering, sequence length filtering, sequence redundancy removal, and structural integrity checks, an initial high-quality set of RNA structures is obtained. Specific filtering criteria can be found in publicly available RNA structure dataset construction methods, and will not be elaborated here.

[0050] To ensure the family representativeness of the dataset, the cmsearch program of the INFERNAL tool was used to match the deredundant sequences with covariance models in the Rfam database. Within sequence families with the same best-matching covariance model, only the sequence with the highest score was retained.

[0051] Following the above screening process, 147 high-quality non-redundant RNA structure data were obtained, each belonging to a different Rfam family. These 147 non-redundant RNA structure data were randomly divided into a training set (111 samples) and an independent test set (TS36, containing 36 samples) at a 3:1 ratio. The test set remained strictly independent and was used only for final performance evaluation after model training was completed; it did not participate in any training or hyperparameter tuning process.

[0052] In this embodiment, the minimum distance between non-hydrogen atoms of nucleotides, 8.0 Å, is used as the threshold to define spatial contact between residues. Nucleotide pairs with a distance not exceeding this threshold are marked as contact (positive samples), and those with a distance greater than this threshold are marked as non-contact (negative samples). A binary true label matrix is ​​constructed in this way for model training and evaluation.

[0053] Step S1: In this embodiment, the RNAcentral database is used as the homology search database. k=3 and step size is 1. The frequency of occurrence of 64 triplets (composed of the combination of four RNA bases: adenine A, guanine G, uracil U and cytosine C) in each sequence is extracted to form a 64-dimensional sequence composition feature vector.

[0054] After normalizing the feature vectors, the k-means clustering algorithm is used to divide the homology search database into multiple independent sequence clusters. In this embodiment, the number of clusters is set to 3, and the sequence data proportions of the three sequence clusters are approximately 24.9%, 51.1%, and 24.0%, respectively. The selection of the number of clusters needs to consider the following factors: if the number of clusters is too small, the fusion of multi-source differential evolution features will be insufficient; if the number of clusters is too large, the channel dimension will increase, leading to a surge in network parameters, which can easily cause overfitting and reduce the efficiency of homology search. Experiments have verified that a cluster number of 3 can achieve the optimal balance between computational efficiency and prediction accuracy. The distance metric between sequence feature vectors is the L2 norm (Euclidean distance), and its calculation formula is as follows:

[0055]

[0056] in, and These represent the normalized frequencies of the two nucleotide sequences on the i-th 3-mer fragment (triple fragment).

[0057] Step S2: Obtain the target RNA sequence, use the cmsearch program to calculate its matching score with all covariance models in the Rfam database, and select the covariance model with the highest score as the guiding model for subsequent homology search.

[0058] Based on the aforementioned guiding model, homologous sequence searches are performed independently within each sequence cluster, with homologous sequences filtered using an E-value threshold. For the retrieved homologous sequences, the `cmalign` function of the `INFERNAL` tool is invoked to perform sequence alignment guided by a covariance model, using the guiding model as a template.

[0059] In the post-alignment processing stage, using the target RNA sequence as a reference, all alignment columns corresponding to gap positions in the target sequence are removed, so that the column dimension of each sub-multiple sequence alignment is consistent with the length L of the target RNA sequence, and three independent sub-multiple sequence alignments are generated, denoted as MSA1, MSA2 and MSA3.

[0060] Step S3: Using pseudo-likelihood maximization direct coupling analysis, MSA1, MSA2, and MSA3 are processed independently, each generating an L×L dimensional evolutionary coupling feature matrix. The three evolutionary coupling feature matrices are concatenated to construct a multi-channel feature tensor of dimension L×L×3, which serves as the input to the hourglass-shaped residual encoder-decoder network in step S4.

[0061] Step S4: The hourglass-shaped residual encoder-decoder network adopts an encoder-decoder architecture to perform feature extraction and contact prediction on the L×L×3 multi-channel feature tensor. The overall network topology is as follows: Figure 1 As shown.

[0062] The network's weight initialization strategy is as follows: all two-dimensional convolutional layers are initialized using the Kaiming normal distribution to maintain the stability of the signal variance during forward propagation; the final single-channel output layer is initialized using the Xavier uniform distribution.

[0063] In this embodiment, the encoder module contains two cascaded residual convolutional blocks, with the number of feature channels progressively increasing from 3 to 4 to 8 to achieve multi-level feature abstraction. Each residual convolutional block consists of a main branch and a residual branch: the main branch uses a double-layer 3×3 convolutional kernel with an equal-size edge padding strategy to expand the receptive field while maintaining the spatial resolution of the feature map. The residual branch achieves linear transformation and dimension matching across channel dimensions through 1×1 convolution. The outputs of the two branches are element-wise added and then passed through the ReLU activation function to form a complete residual mapping, effectively alleviating the gradient vanishing problem in deep networks. To prevent overfitting, a two-dimensional random deactivation layer with a dropout rate of 0.5 is introduced after the encoder output in this embodiment.

[0064] In this embodiment, the decoder module includes two cascaded residual convolutional blocks symmetrical to the encoder. The feature channels are progressively compressed in an 8→4→2 manner to achieve feature reconstruction and spatial resolution restoration. LeakyReLU is uniformly used as the non-linear activation function throughout the network to prevent neuron deactivation.

[0065] In this embodiment, the multi-channel feature map output by the decoder is reduced to a single channel via 1×1 convolution and mapped to the [0, 1] probability interval using a Sigmoid activation function to generate an initial RNA residue contact probability matrix. Subsequently, a symmetry post-processing is performed: for the predicted probability values ​​of two symmetrical positions (i, j) and (j, i) representing the same nucleotide pair in the matrix, the larger of the two values ​​is taken to update the corresponding two positions in the matrix, generating a symmetrical final RNA residue contact probability matrix. This ensures that the prediction results conform to the undirected symmetry constraint of spatial interactions between RNA molecule residues.

[0066] In this embodiment, model training is based on a GPU computing platform supporting CUDA acceleration and the PyTorch deep learning framework. The batch size is set to 1 during training to accommodate the dynamic input dimensions of RNA molecules of different lengths.

[0067] In this embodiment, the Adam optimizer is used for parameter updates, with the initial learning rate set to 0.001 and the weight decay factor set to 1× Simultaneously, an adaptive learning rate scheduling strategy based on validation set loss is introduced. When no decrease in validation set loss is observed for five consecutive training epochs, the learning rate is adjusted with a decay factor of 0.5, and the minimum learning rate is limited to 1× .

[0068] To address the extreme imbalance between positive and negative samples in RNA residue contact prediction, this invention constructs an absolute error loss function, the mathematical expression of which is:

[0069]

[0070] Where N is the total number of valid elements in the matrix; The predicted contact probability of residue pair (i, j) output by the network; ∈{0,1} represents a true binary contact label. The asymmetric penalty weight is defined as follows:

[0071]

[0072] in, The penalty weight hyperparameter is greater than 1 for positive samples. This loss function guides the network to focus on learning the spatial contact features of positive samples by imposing a higher penalty weight on real contact samples.

[0073] In this embodiment, after the model outputs the final symmetric contact probability matrix, the continuous probability values ​​need to be binarized into contact determination results. After a full threshold scan on the validation set, based on the principle of maximizing the F1 score, the contact determination threshold is set to 0.4. That is, when the predicted probability is not lower than 0.4, the nucleotide pair is determined to be in spatial contact; otherwise, it is determined to be non-contact.

[0074] To objectively measure the predictive performance of the model, this embodiment uses precision, recall, F1 score, and Matthews correlation coefficient (MCC) as core evaluation indicators. Their mathematical definitions follow well-known formulas and will not be elaborated here.

[0075] On an independent test set containing 36 non-redundant RNA structures, prediction performance was graded and evaluated based on sequence spacing distance (|i−j|), and the results are as follows: Figure 2 As shown. The method in this embodiment is applicable across the entire distance range ( >0), medium- to long-range contact ( >4) and remote contact ( >24) The evaluation results for the three distance intervals are as follows:

[0076] In the full range ( Within the range of >0, Precision is 0.848, Recall is 0.611, F1 is 0.692, and MCC is 0.682; in medium-to-long-range contact ( Within >4), Precision is 0.724, Recall is 0.480, F1 is 0.551, and MCC is 0.546; in remote contact ( Within >24), Precision was 0.700, Recall was 0.417, F1 was 0.499, and MCC was 0.503. These results demonstrate that the method of this invention possesses high prediction accuracy across the entire distance range, maintaining a high level of accuracy even for medium- to long-range contacts, thus providing spatial constraint information for RNA three-dimensional structure modeling.

[0077] To verify the effectiveness of the multi-channel input strategy, this embodiment compares the single MSA strategy with the MSAs strategy. The single MSA strategy directly searches for homologous sequences in a complete homology search database without clustering, constructs a single multiple sequence alignment, and extracts the evolutionary coupling feature matrix as input, with an input dimension of L×L×1. The MSAs strategy, on the other hand, first performs pre-clustering of the database based on sequence composition features, independently searches for homologous sequences in each sequence cluster, constructs sub-multiple sequence alignments, and uses the evolutionary coupling feature matrix of each sub-multiple sequence alignment as an independent channel input, with an input dimension of L×L×3. The comparison results of the two strategies are as follows: Figure 3 As shown.

[0078] like Figure 3 As shown, on the independent test set TS36, the MSAs policy performs well across the entire range ( >0), medium-to-long-range contact ( >4) and remote contact ( >24) Across the three intervals, the MSAs strategy comprehensively outperformed the single MSA strategy in all four core metrics: Precision, Recall, F1 score, and MCC. Taking medium- to long-range contact prediction as an example, the MSAs strategy improved the F1 score from 0.551 to 0.645 and the MCC from 0.546 to 0.634. These results demonstrate that by performing pre-clustering of the sequence database based on sequence composition features and constructing a multi-channel evolutionary coupling feature tensor, compared to directly constructing a single multiple sequence alignment in the complete database, the differentiated evolutionary patterns between sequence clusters can be preserved, thus improving the accuracy of RNA residue contact prediction.

[0079] Figure 4 This demonstrates the effectiveness of the method of the present invention in predicting contacts of RNA structure 1Y26 chain A (71 nucleotides in length).

[0080] like Figure 4 As shown in Figure A, the red lines represent base pairing relationships, and the yellow lines represent long-range tertiary structure spatial contacts. The method of this invention has an average prediction probability of 0.919 for 22 base pairings, indicating a high recognition ability for Watson-Crick pairings.

[0081] like Figure 4 As shown in B in the diagram, the red spiral region in the tertiary structure visualization corresponds to the stem structure predicted with high probability, forming a stable double helix conformation in three-dimensional space, consistent with the prediction results.

[0082] like Figure 4 As shown in Figure C, the upper triangle represents the RNA residue contact probability matrix output by the method of this invention, and the lower triangle represents the actual contact tags. For residue pairs with large sequence intervals (residue pairs 10-41, 25-49, and 26-48, with sequence intervals of 31, 24, and 22 respectively), the corresponding average predicted probability is 0.45. The above prediction results intuitively demonstrate the probability prediction capability of the method of this invention for RNA residue contact.

[0083] Another invention, embodiments of the present invention, also provide an RNA residue contact prediction system based on an hourglass-shaped residual encoder-decoder network, comprising:

[0084] The sequence clustering module is used to vectorize the nucleotide sequences in the RNA sequence database based on the sequence composition features of RNA sequences, and then divide the RNA sequence database into multiple independent sequence clusters with low redundancy through a clustering algorithm.

[0085] The homology search module is used to obtain the target RNA sequence, calculate and determine the optimal covariance model of the target RNA sequence in the Rfam known RNA family database, and use the covariance model as a guide to perform homology sequence search in each independent sequence cluster obtained by the sequence clustering module, and construct the sub-multiple sequence alignment corresponding to each independent sequence cluster.

[0086] The coupling feature extraction module is used to perform direct coupling analysis on each sub-sequence alignment and extract the corresponding evolutionary coupling feature matrix.

[0087] The multi-channel fusion and prediction module is used to concatenate multiple evolutionary coupling feature matrices into a multi-channel input feature tensor, which is then input into an hourglass-shaped residual encoder-decoder network for feature fusion and contact prediction. After symmetry processing, the module outputs the RNA residue contact probability matrix of the target RNA sequence.

[0088] The specific implementation methods of each module are the same as those of each step, and will not be described in the embodiments of this invention.

[0089] Furthermore, this invention also provides an RNA residue contact prediction device based on an hourglass-shaped residual codec network, including a memory and a processor. The memory is used to store a computer program, and the processor executes the above-mentioned RNA residue contact prediction method based on an hourglass-shaped residual codec network when running the computer program.

[0090] It should be understood that the above description of the embodiments is quite detailed and should not be construed as limiting the scope of protection of this invention. Those skilled in the art, guided by the teachings of this invention, can make substitutions or modifications without departing from the scope of protection of the claims, all of which fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for predicting RNA residue contacts based on an hourglass-shaped residual encoder-decoder network, characterized in that, Includes the following steps: S1: Based on the sequence composition features of RNA sequences, the nucleotide sequences in the RNA sequence database are vectorized and then the RNA sequence database is divided into multiple independent sequence clusters with low redundancy through a clustering algorithm. S2: Obtain the target RNA sequence, calculate and determine the optimal covariance model of the target RNA sequence in the Rfam known RNA family database, and use the covariance model as a guide to perform homologous sequence search in each independent sequence cluster obtained in step S1, and construct the sub-multiple sequence alignment corresponding to each independent sequence cluster. S3: Perform direct coupling analysis on each sub-sequence alignment and extract the corresponding evolutionary coupling feature matrix; S4: Concatenate multiple evolutionary coupling feature matrices into a multi-channel input feature tensor, input it into an hourglass-shaped residual encoder-decoder network for feature fusion and contact prediction, and output the RNA residue contact probability matrix of the target RNA sequence after symmetry processing.

2. The RNA residue contact prediction method based on an hourglass-shaped residual encoder-decoder network according to claim 1, characterized in that: The vectorized representation based on sequence composition features described in step S1 constructs a fixed-length feature vector by statistically analyzing the occurrence frequency of each triplet fragment in the nucleotide sequence; the clustering algorithm is based on the L2 norm distance metric between feature vectors, grouping sequences with similar composition features into the same sequence cluster.

3. The RNA residue contact prediction method based on an hourglass-shaped residual encoder-decoder network according to claim 1, characterized in that: In step S2, the covariance model is determined and the sub-multiple sequence alignment is constructed, specifically including: The matching scores of the target RNA sequence with each covariance model in the Rfam family database are calculated, and the covariance model with the highest score is selected as the guiding model for homology search. Based on the guiding model, homology sequence search is performed independently in each sequence cluster, and an E-value threshold is set to filter homology sequences. For the retrieved homology sequences, sequence alignment guided by the covariance model is performed with the target RNA sequence as a reference. All alignment columns corresponding to the gap positions in the target sequence are pruned so that the column dimension of each sub-multiple sequence alignment is consistent with the length of the target RNA sequence, and sub-multiple sequence alignments corresponding to each sequence cluster are generated respectively.

4. The RNA residue contact prediction method based on an hourglass-shaped residual encoder-decoder network according to claim 1, characterized in that: The direct coupling analysis method described in step S3 is an analysis method for extracting direct evolutionary coupling relationships between residues in multiple sequence alignments, including a pseudo-likelihood maximization direct coupling analysis method; by processing each of the sub-multiple sequence alignments independently, corresponding... 3D evolutionary coupling feature matrix; steps The multi-channel input feature tensor described in S4 consists of multiple Constructed by splicing together evolutionary coupling feature matrices, with dimension 1. Where N is the number of sequence clusters, The length of the target RNA sequence is nucleotides.

5. The RNA residue contact prediction method based on an hourglass-shaped residual encoder-decoder network according to claim 1, characterized in that: The hourglass-shaped residual codec network architecture described in step S4 specifically includes: The encoder module consists of cascaded residual convolutional blocks, which perform multi-level feature abstraction on the multi-channel feature tensor. The number of feature channels gradually increases along the network depth direction, and outputs a high-dimensional abstract feature map. Decoder module: Composed of cascaded residual convolutional blocks symmetrical to the encoder module, it performs feature reconstruction and spatial resolution restoration on the high-dimensional abstract feature map. The number of feature channels is progressively compressed along the network depth direction to realize feature reconstruction and prediction of contact probability between residues. Single-channel output module: The multi-channel feature map output by the decoder module is reduced to a single channel through convolution, and the predicted value is mapped to the [0,1] probability interval by the Sigmoid activation function, and the contact probability matrix between RNA residues of the target RNA sequence is output.

6. The RNA residue contact prediction method based on an hourglass-shaped residual encoder-decoder network according to claim 5, characterized in that: The encoder module contains two cascaded residual convolutional blocks, each consisting of a main branch and a residual branch: the main branch uses a double-layer 3×3 convolutional kernel with an equal-size edge padding strategy to expand the receptive field while maintaining the feature map spatial resolution; the residual branch achieves linear transformation and dimension matching across channels through 1×1 convolution; the outputs of the two branches are added element-wise and then activated by the LeakyReLU activation function to form a complete residual mapping, which is then introduced into a two-dimensional random deactivation layer after the encoder module output; The weight initialization strategy of the hourglass-shaped residual codec network is as follows: all two-dimensional convolutional layers are initialized with a Kaiming normal distribution to maintain the stability of the signal variance during forward propagation; the final single-channel output layer is initialized with a Xavier uniform distribution.

7. The RNA residue contact prediction method based on an hourglass-shaped residual encoder-decoder network according to claim 1, characterized in that: The hourglass-shaped residual encoder-decoder network is trained using an absolute error loss function. The mathematical expression of the absolute error loss function is as follows: ； Where N is the total number of valid elements in the matrix; The predicted contact probability of residue pair (i, j) output by the network; ∈{0,1} represents a true binary contact label; The asymmetric penalty weight is defined as follows: ； in, The penalty weight hyperparameter for positive samples greater than 1.

8. The RNA residue contact prediction method based on an hourglass-shaped residual encoder-decoder network according to claim 1, characterized in that: After outputting the RNA residue contact probability matrix in step S4, a symmetry processing step is also included, specifically: For the predicted probability values ​​of the two symmetrical positions (i, j) and (j, i) representing the same nucleotide pair (i, j) in the contact probability matrix, the larger of the two values ​​is taken as the final spatial contact probability of the nucleotide pair, and the corresponding two symmetrical positions in the matrix are updated accordingly to generate a symmetrical final contact probability matrix. The symmetry processing eliminates the directional prediction bias generated during unidirectional convolution inference by integrating the prediction results of symmetrical positions in the matrix, so that the output matrix conforms to the undirected symmetry constraint of spatial interaction between RNA residues.

9. The RNA residue contact prediction method based on an hourglass-shaped residual encoder-decoder network according to claim 1, characterized in that: After outputting the final contact probability matrix, the continuous probability values ​​are binarized into contact determination results. Based on maximizing precision, recall, F1 score, or Hughes correlation coefficient, the contact determination threshold is set to [value missing]. That is, when the predicted probability is not lower than If the nucleotide pair is in spatial contact, it is determined that there is contact; otherwise, it is determined that there is no contact.

10. An RNA residue contact prediction system based on an hourglass-shaped residual encoder-decoder network, characterized in that, include: The sequence clustering module is used to vectorize the nucleotide sequences in the RNA sequence database based on the sequence composition features of RNA sequences, and then divide the RNA sequence database into multiple independent sequence clusters with low redundancy through a clustering algorithm. The homology search module is used to obtain the target RNA sequence, calculate and determine the optimal covariance model of the target RNA sequence in the Rfam known RNA family database, and use the covariance model as a guide to perform homology sequence search in each independent sequence cluster obtained by the sequence clustering module, and construct the sub-multiple sequence alignment corresponding to each independent sequence cluster. The coupling feature extraction module is used to perform direct coupling analysis on each sub-sequence alignment and extract the corresponding evolutionary coupling feature matrix. The multi-channel fusion and prediction module is used to concatenate multiple evolutionary coupling feature matrices into a multi-channel input feature tensor, which is then input into an hourglass-shaped residual encoder-decoder network for feature fusion and contact prediction. After symmetry processing, the module outputs the RNA residue contact probability matrix of the target RNA sequence.

Images