RNA-small molecule binding site prediction method and system

Abstract

The present invention discloses a method and system for predicting RNA-small molecule binding sites. The method comprises obtaining global information and local information based on a motif of a full-length RNA and a motif of its local neighboring RNAs; inputting the global information into a two-dimensional feature fusion network to obtain global features, wherein the two-dimensional feature fusion network includes a convolutional neural network channel and an attention network channel; inputting the local information into a fully connected neural network to obtain local features; inputting the fused feature information after integrating the global features and the local features into an Inception module of a one-dimensional feature fusion for deep fusion to obtain deep fusion features; and inputting the deep fusion features into an output layer to predict the probability value of the binding site. The present invention greatly improves the effective utilization of RNA feature information and reduces the number of parameters, so that the network can maintain the prediction performance of RNA-small molecule binding sites while also being lightweight.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method and system for predicting RNA-small molecule binding sites. Background Art

[0002] RNA plays an essential role in many biochemical processes, such as gene expression regulation, protein synthesis, and signal transduction, making it an indispensable macromolecule in living organisms. Some RNAs in living organisms can bind to DNA or protein targets, thereby inhibiting or promoting biochemical reactions. In recent years, RNA has shown great potential as a therapeutic target for small molecule binding. However, due to the considerable flexibility of RNA structure, identifying stable RNA folding conformations is a challenging task. Therefore, to better understand the recognition mechanisms of different small molecules binding to RNA, predicting RNA-small molecule binding sites has become a pressing task. To date, only a few computational methods have been developed for predicting RNA-ligand binding sites, including Rsite, Rsite2, RBind, RNAsite, and RLBind. Rsite is a computational method that uses tertiary structure information to identify functional sites in RNA molecules. Subsequently, Rsite2 evolved from Rsite, shifting its computational approach from tertiary structure to secondary structure. RBind primarily relies on RNA structural information to predict RNA-small molecule binding sites. RNAsite, a state-of-the-art machine learning-based method, utilizes both RNA sequence and structural information to predict RNA-small molecule binding sites. RLBind, a state-of-the-art RNA binding site prediction method based on convolutional neural networks, also leverages RNA sequence and structure information. However, Rsite, Rsite2, and RBind only use RNA structure information to predict RNA-ligand binding sites, without utilizing RNA sequence information. While RNAsite and RLBind utilize both RNA structure and sequence information, they only calculate the structure and sequence information separately and then simply add them together, failing to effectively integrate the RNA structure and sequence information.

[0003] In the related art, patent application documents with publication numbers CN111798921A, CN112863597A and CN113241117A all solve the problem of RNA and protein binding sites, but proteins are biological macromolecules rather than small molecules, and no feature fusion is performed. Summary of the Invention

[0004] The technical problem to be solved by the present invention is how to make the network lightweight while maintaining the prediction performance of RNA-small molecule binding sites.

[0005] The present invention solves the above technical problems through the following technical means:

[0006] The present invention proposes a method for predicting RNA-small molecule binding sites, the method comprising:

[0007] Based on the motifs of the full-length RNA and the motifs of its local neighboring RNAs, global information and local information are obtained, wherein the global information includes the sequence and structural feature information of the RNA itself, and the local information includes the sequence and structural feature information of the local neighboring RNAs;

[0008] Sending the global information into a two-dimensional feature fusion network to obtain global features, wherein the two-dimensional feature fusion network includes a convolutional neural network channel and an attention network channel;

[0009] Sending the local information into a fully connected neural network to obtain local features;

[0010] The fused feature information after integrating the global features and the local features is sent to the Inception module of one-dimensional feature fusion for deep fusion to obtain deep fusion features;

[0011] The deep fusion features are sent to the output layer to predict the probability value of the binding site.

[0012] Furthermore, the motif based on the full-length RNA and the motif of its local neighbor RNAs is used to obtain global information and local information, including:

[0013] Using the motifs of the full-length RNA and its local neighbor RNAs, the input properties are calculated;

[0014] Encoding the input attributes to obtain a multi-dimensional vector, wherein the vector information includes the sequence-dependent attributes and structure-dependent attributes of the RNA itself and the sequence-dependent attributes and structure-dependent attributes of local neighboring RNAs;

[0015] Obtaining the global information based on the sequence-dependent properties and structure-dependent properties of the RNA itself;

[0016] The local information is obtained based on the sequence-dependent properties and structure-dependent properties of the local neighbor RNAs.

[0017] Furthermore, encoding the input attributes to obtain a multi-dimensional vector includes:

[0018] The input attributes are encoded from sequence-related attributes, structure-related attributes, network topology attributes and accessible surface areas to obtain a ten-dimensional vector.

[0019] Furthermore, the sequence-dependent attributes include nucleotide type and evolutionary conservation score;

[0020] The structure-dependent properties include network topology properties, biochemical properties and ASAs.

[0021] Furthermore, the convolutional neural network channel includes a two-dimensional convolution layer and a pooling layer, the input of the two-dimensional convolution layer is the global information, and the output of the two-dimensional convolution layer is connected to the pooling layer;

[0022] The attention network channel includes a bidirectional long short-term memory network BiLSTM and an attention network, the input of the BiLSTM is the global information, and the output of the BiLSTM is connected to the attention network;

[0023] The global features are obtained by concatenating the output features of the pooling layer and the output features of the attention network.

[0024] Furthermore, the BiLSTM includes an input layer, a hidden layer, and an output layer connected in sequence, and the attention network includes a network fusion layer Multiply_1, a softmax layer, and a network fusion layer Multiply_2;

[0025] The hidden layer and output layer of the BiLSTM are both connected to the network fusion layer Multiply_1, the output of the network fusion layer Multiply_1 is connected to the softmax layer, and the output of the softmax layer and the output layer of the BiLSTM are both connected to the network fusion layer Multiply_2.

[0026] Furthermore, the step of feeding the local information into a fully connected neural network to obtain local features includes:

[0027] The local information is sent to the fully connected neural network for dimensionality reduction processing to obtain the local features.

[0028] Furthermore, the Inception module includes one-dimensional convolution kernels of different sizes, each of the one-dimensional convolution kernels is arranged in parallel, and each of the one-dimensional convolution kernels is followed by a batch normalization layer.

[0029] Furthermore, the Inception module includes one-dimensional convolution kernels Conv_1, Conv_2, Conv_3, Conv_4, Conv_5, Conv_6, Conv_8, Conv_8, pooling layer Pooling_2 and concatenation layer Concatenate;

[0030] The input of the one-dimensional convolution kernel Conv_1 is the fused feature information, and the output of the one-dimensional convolution kernel Conv_1 is connected to the concatenation layer Concatenate through the batch normalization layer BN_1;

[0031] The input of the one-dimensional convolution kernel Conv_2 is connected to the pooling layer Pooling_2, and the output is connected to the concatenation layer Concatenate through the batch normalization layer BN_2. The input of the pooling layer Pooling_2 is the fused feature information;

[0032] The input of the one-dimensional convolution kernel Conv_3 is the fused feature information, the output of the one-dimensional convolution kernel Conv_3 is connected to the one-dimensional convolution kernel Conv_4 through the batch normalization layer BN_3, and the output of the one-dimensional convolution kernel Conv_4 is connected to the concatenation layer Concatenate through the batch normalization layer BN_4;

[0033] The input of the one-dimensional convolution kernel Conv_5 is the fused feature information. The output of the one-dimensional convolution kernel Conv_5 is connected to the one-dimensional convolution kernel Conv_6 via the batch normalization layer BN_5. The output of the one-dimensional convolution kernel Conv_6 is connected to the one-dimensional convolution kernel Conv_7 via the batch normalization layer BN_6. The output of the one-dimensional convolution kernel Conv_7 is connected to the concatenation layer Concatenate via the batch normalization layer BN_7.

[0034] The output of the concatenation layer Concatenate is connected to the batch normalization layer BN_8 through the one-dimensional convolution kernel Conv_8.

[0035] In addition, the present invention also proposes an RNA-small molecule binding site prediction system, which includes:

[0036] A feature encoding module for obtaining global information and local information based on the motifs of the full-length RNA and the motifs of its local neighboring RNAs, wherein the global information includes the sequence and structural feature information of the RNA itself, and the local information includes the sequence and structural feature information of the local neighboring RNAs;

[0037] The prediction module is used to input the global information into a two-dimensional feature fusion network to obtain global features, wherein the two-dimensional feature fusion network includes a convolutional neural network channel and an attention network channel; input the local information into a fully connected neural network to obtain local features; input the fused feature information after integrating the global features and the local features into an Inception module of a one-dimensional feature fusion for deep fusion to obtain deep fusion features; and input the deep fusion features into an output layer to predict the probability value of the binding site.

[0038] The advantages of the present invention are:

[0039] (1) The present invention utilizes the sequence and structural features of RNA itself as global information, and a limited number of adjacent nucleotide sequences and structural features as local information, and fuses these two types of information through a multidimensional feature fusion network, thereby improving the probability of predicting RNA-small molecule binding sites; a two-dimensional feature fusion network is used to perform two-dimensional feature fusion on the global information, and then a one-dimensional feature network is used to perform one-dimensional feature fusion on the local information. These two fusions of features in different dimensions greatly improve the effective utilization of RNA feature information and further reduce the number of parameters, so that the network can maintain the prediction performance of RNA-small molecule binding sites while also taking into account lightweight.

[0040] (2) The two-dimensional feature fusion network consists of two independent data processing channels: a convolutional neural network channel and an attention network channel. It can effectively extract global information features from different angles without interfering with each other. The original global information is added after passing through these two data processing channels to obtain a new global information expression.

[0041] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 1 is a flow chart of a method for predicting RNA-small molecule binding sites according to an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the principle of RNA-small molecule binding site prediction in one embodiment of the present invention;

[0044] Figure 3 1 is a schematic diagram of the structure of a multi-dimensional feature fusion network in one embodiment of the present invention;

[0045] Figure 4 2 is a schematic diagram of the structure of the Inception module in one embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of a data set in an embodiment of the present invention;

[0047] Figure 6 Schematic diagram of the structure of an RNA-small molecule binding site prediction system proposed in one embodiment of the present invention. DETAILED DESCRIPTION

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0049] like Figures 1 to 2 As shown, the first embodiment of the present invention discloses a method for predicting RNA-small molecule binding sites, the method comprising the following steps:

[0050] S10, obtaining global information and local information based on the motif of the full-length RNA and the motif of its local neighboring RNAs, wherein the global information includes sequence and structural feature information of the RNA itself, and the local information includes sequence and structural feature information of the local neighboring RNAs;

[0051] S20, sending the global information into a two-dimensional feature fusion network to obtain global features, where the two-dimensional feature fusion network includes a convolutional neural network channel and an attention network channel;

[0052] S30, sending the local information into a fully connected neural network to obtain local features;

[0053] S40, sending the fusion feature information obtained by integrating the global features and the local features into the one-dimensional feature fusion Inception module for deep fusion to obtain deep fusion features;

[0054] S50: Send the deep fusion features to the output layer to predict the probability value of the binding site.

[0055] This embodiment uses the sequence and structural features of RNA itself as global information, and a limited number of adjacent nucleotide sequences and structural features as local information. These two types of information are fused through a multidimensional feature fusion network, thereby improving the probability of predicting RNA-small molecule binding sites. The multidimensional feature fusion network here includes a two-dimensional feature fusion network and a fully connected neural network. The global features output by the two-dimensional feature fusion network and the local features output by the fully connected network are connected to obtain a new two-dimensional feature expression containing global and local information. The new two-dimensional feature information is then reduced in dimension and input into the Inception module, which can compress and fuse the two-dimensional feature information into one-dimensional feature information. A three-layer fully connected neural network is used as the final output layer. The deeply fused features after passing through the Inception module are fed into the output layer, and the output of this output layer is used as the probability value for predicting the binding site.

[0056] Specifically, the two-dimensional feature fusion network includes a convolutional neural network channel and an attention network channel. When performing two-dimensional feature fusion on global information, the sequence and structural features of RNA itself can be extracted more efficiently; and then one-dimensional feature fusion is performed on local information. These two feature fusions of different dimensions can make full use of the sequence and structural features of RNA, thereby better predicting the binding sites of RNA and small molecules, and further reducing the number of parameters, so that the network can maintain the prediction performance of RNA-small molecule binding sites while also taking into account lightweight.

[0057] In one embodiment, step S10: obtaining global information and local information based on the motif of the full-length RNA and the motifs of its local neighboring RNAs, comprises the following steps:

[0058] S11. Calculate input attributes using the motifs of the full-length RNA and its local neighbor RNAs;

[0059] S12. Encode the input attributes to obtain a multi-dimensional vector, where the vector information includes the sequence-dependent attributes and structure-dependent attributes of the RNA itself and the sequence-dependent attributes and structure-dependent attributes of local neighboring RNAs;

[0060] S13, obtaining the global information based on the sequence-dependent properties and structure-dependent properties of the RNA itself;

[0061] S14. Obtain the local information based on the sequence-dependent properties and structure-dependent properties of the local neighbor RNA.

[0062] Specifically, for the construction of RNA global features, this embodiment adopts a fixed measure for RNAs of different sequence lengths, fixing the length to 64. RNA sequences exceeding this fixed length are truncated, and RNAs with a length not exceeding 64 are padded with zeros.

[0063] In one embodiment, the step S12 of encoding the input attributes to obtain a multi-dimensional vector specifically includes:

[0064] The input attributes are encoded from sequence-related attributes, structure-related attributes, network topology attributes and accessible surface areas to obtain a ten-dimensional vector.

[0065] In one embodiment, the sequence-dependent attributes include nucleotide type and evolutionary conservation score; the structure-dependent attributes include network topology, biochemical properties, and ASAs.

[0066] Specifically, this embodiment encodes features mainly from four aspects, including: sequence-related attributes, structure-related attributes, network topology attributes, and accessible surface area.

[0067] (1) Sequence-related attributes are represented by a 5-dimensional vector, which includes a 4-dimensional one-hot encoding for nucleotide type and a 1-dimensional one-hot encoding for evolutionary conservation.

[0068] (2) Three types of structure-related properties were used: two-dimensional vector-encoded network topological properties, two-dimensional vector-encoded biochemical properties, and one-dimensional vector-encoded ASA. The two-dimensional vector-encoded biochemical properties included molecular weight and side chain pKa.

[0069] (3) For network topology, the RNA tertiary structure was used to construct a network, in which each nucleotide was represented as a node and the non-covalent interaction between a pair of heavy atoms of two non-contiguous nucleotides was defined as an edge. Then, two network topology properties, namely proximity and degree, were calculated to evaluate the importance of each node.

[0070] (4) In terms of accessible surface area, the solvent accessible surface area (ASA) of nucleotides can provide key information for describing functional sites or core structural regions. ASA is defined as reflecting the degree to which a nucleotide is exposed to the solvent. Specifically, nucleotides with higher ASAs are usually located on the surface of the RNA structure, while nucleotides with lower ASAs are usually located inside the binding cavity. Therefore, RNAsol was used in this example to calculate the ASA of nucleotides.

[0071] In one embodiment, if Figure 3 As shown, the convolutional neural network channel includes a two-dimensional convolution layer Convolution and a pooling layer Pooling_1, the input of the two-dimensional convolution layer Convolution is the global information, and the output of the two-dimensional convolution layer Convolution is connected to the pooling layer Pooling_1;

[0072] The attention network channel includes a bidirectional long short-term memory network BiLSTM and an attention network Attention, the input of the BiLSTM is the global information, and the output of the BiLSTM is connected to the attention network Attention;

[0073] The global features are obtained by concatenating the output features of the pooling layer Pooling_1 and the output features of the attention network Attention.

[0074] In one embodiment, the BiLSTM includes an input layer, a hidden layer, and an output layer connected in sequence, and the attention network Attention includes a network fusion layer Multiply_1, a softmax layer, and a network fusion layer Multiply_2;

[0075] The hidden layer and output layer of the BiLSTM are both connected to the network fusion layer Multiply_1, the output of the network fusion layer Multiply_1 is connected to the softmax layer, and the output of the softmax layer and the output layer of the BiLSTM are both connected to the network fusion layer Multiply_2.

[0076] Specifically, this embodiment performs a dot multiplication on the BiLSTM hidden layer and the output layer, then performs dimensionality reduction on the product after the dot multiplication, and then uses the reduced product as a weight matrix after processing with a softmax function, and then multiplies the output result of the BiLSTM by the weight matrix as the final output of the branch.

[0077] In one embodiment, the step of feeding the local information into a fully connected neural network to obtain local features includes:

[0078] The local information is sent to the fully connected neural network for dimensionality reduction processing to obtain the local features.

[0079] For local information, this embodiment adopts a dimensionality reduction method to compress the data space in the spatial dimension, so that the data in the local information is more compact.

[0080] In one embodiment, the Inception module includes one-dimensional convolution kernels of different sizes, each of the one-dimensional convolution kernels is arranged in parallel, and each of the one-dimensional convolution kernels is followed by a batch normalization layer.

[0081] Furthermore, if Figure 4 As shown, the Inception module includes one-dimensional convolution kernels Conv_1, Conv_2, Conv_3, Conv_4, Conv_5, Conv_6, Conv_8, Conv_8, pooling layer Pooling_2 and concatenation layer Concatenate;

[0082] The input of the one-dimensional convolution kernel Conv_1 is the fused feature information, and the output of the one-dimensional convolution kernel Conv_1 is connected to the concatenation layer Concatenate through the batch normalization layer BN_1;

[0083] The input of the one-dimensional convolution kernel Conv_2 is connected to the pooling layer Pooling_2, and the output is connected to the concatenation layer Concatenate through the batch normalization layer BN_2. The input of the pooling layer Pooling_2 is the fused feature information;

[0084] The input of the one-dimensional convolution kernel Conv_3 is the fused feature information, the output of the one-dimensional convolution kernel Conv_3 is connected to the one-dimensional convolution kernel Conv_4 through the batch normalization layer BN_3, and the output of the one-dimensional convolution kernel Conv_4 is connected to the concatenation layer Concatenate through the batch normalization layer BN_4;

[0085] The input of the one-dimensional convolution kernel Conv_5 is the fused feature information. The output of the one-dimensional convolution kernel Conv_5 is connected to the one-dimensional convolution kernel Conv_6 via the batch normalization layer BN_5. The output of the one-dimensional convolution kernel Conv_6 is connected to the one-dimensional convolution kernel Conv_7 via the batch normalization layer BN_6. The output of the one-dimensional convolution kernel Conv_7 is connected to the concatenation layer Concatenate via the batch normalization layer BN_7.

[0086] The output of the concatenation layer Concatenate is connected to the batch normalization layer BN_8 through the one-dimensional convolution kernel Conv_8.

[0087] Specifically, the Inception module has four branches, all of which are composed of one-dimensional convolution kernels of different sizes. Finally, the results of the four branches are spliced ​​together, and then the number of channels is compressed to 3 through a one-dimensional convolution kernel of size 1x1, which is used as the final output of the Inception module.

[0088] In the embodiment, the new two-dimensional feature information is reduced in dimension and then input into a one-dimensional feature fusion module composed of one-dimensional convolution, which can compress and fuse the two-dimensional feature information into one-dimensional feature information.

[0089] Furthermore, a simple three-layer fully connected neural network is used as the output layer. The number of neurons in the three fully connected layers is 196, 96, and 1. The output of the last fully connected layer is passed through the sigmoid activation function to predict the probability of whether the site is an RNA-small molecule binding site.

[0090] It should be noted that this embodiment pre-constructs a site prediction model, which includes the above-mentioned two-dimensional feature fusion network, fully connected neural network, Inception module and output layer. The encoded global information and local information are respectively sent to the two-dimensional feature fusion network and the fully connected neural network. The output of the two-dimensional feature fusion network and the output of the fully connected neural network are integrated as the input of the Inception module, and the output of the Inception module is connected to the output layer.

[0091] After the above-mentioned site prediction model is constructed, the site prediction model is trained using the constructed training set, and the trained site prediction model is used to perform site prediction on the current RNA motif to be processed.

[0092] Specifically, the site prediction model was run on an Ubuntu system with an Nvidia TitanX GPU and 64GB of memory. In this network, to avoid the problem of vanishing gradients, the ReLU activation function was used, which is a common and effective activation function. At the same time, to avoid overfitting, the model used a dropout of 0.3. The adaptive optimizer Adam was used to optimize the parameterized model, and the learning rate was set to 0.0001. For the choice of loss function, the binary cross entropy (BCELoss) was used. This is a loss function that is very common in binary classification tasks and can effectively reduce the gap between the target value and the predicted value. The specific training process includes:

[0093] (1) Dataset construction: The training and test data in TEST18 and TEST9 were obtained from the RNAsite and Rbind methods, respectively. Specifically:

[0094] The TEST18 dataset was collected from the Protein Data Bank (PDB). It contains 78 RNA-ligand complexes, based on the PDB IDs provided by the RNAsite method. In this dataset, if an RNA molecule can form a complex with one or more ligands, the complex is selected for binding site prediction. Ligands in this context refer to small molecules with low molecular weight, excluding water molecules. RNA monomers were then separated from the complexes for cluster analysis and feature calculation. According to RNAsite, we removed 634 redundant structures through pairwise structural similarity clustering. Sequence-based similarity clustering was then used to eliminate redundancy between the training and test sets. The 78 RNAs were then divided into 57 clusters with a sequence similarity of 30%. The dataset was divided into training, validation, and test sets. The training and validation sets contained 42 distinct clusters and 60 RNAs (T60). These 60 RNAs were randomly divided into the training and validation sets at a 9:1 ratio. The test set contained 15 distinct clusters and 18 RNAs (T18).

[0095] The TEST9 dataset, obtained from the RBind method, contains 22 RNAs. During the RNA tertiary structure modeling process, RNA structures with multi-stranded or pseudoknot interactions were removed, retaining only 19 of these structures. Furthermore, due to different data collection sources, 10 RNAs in the existing training set duplicated the 19 RNAs in this dataset. To better evaluate model performance, we removed these 10 duplicates from the dataset, retaining only 9 (T9). This dataset served as the test set.

[0096] (2) The three datasets (T60, T18 and T9) introduced in the dataset section are used. Figure 5 The original data is encoded according to the feature encoding method, which is consistent with RLBind. At the same time, RLBind also provides the encoded data. The original data part is shown in Table 1, with the ID and link corresponding to the PDB website.

[0097] (3) The encoded features are fed into the site prediction model for training. The two-dimensional feature fusion module uses two branches. The first branch consists of two modules: convolution and pooling. The convolution module uses a convolution kernel size of 17x10 and a channel number of 128; the pooling kernel size used in the pooling is 64x1. The second branch consists of a BiLSTM and an attention module in series. The middle hidden layer of the BiLSTM is set to 20, the input layer is set to 10, and the output layer is set to 128. In the attention module, the BiLSTM hidden layer and the output layer are multiplied by points, and the product after the point multiplication is then reduced in dimension. The product after the dimension reduction is then processed by the softmax function and used as a weight matrix. The output of the BiLSTM is then multiplied by the weight matrix as the final output of the branch.

[0098] The Inception module has four branches, all of which are composed of one-dimensional convolution kernels of different sizes. Finally, the results of the four branches are spliced ​​together, and then the number of channels is compressed to 3 through a one-dimensional convolution kernel of size 1x1 as the final output of the module.

[0099] The fully connected layer uses a simple three-layer fully connected neural network with 196, 96, and 1 neurons. The output of the last fully connected layer is activated by a sigmoid function to predict the probability of whether the site is an RNA-small molecule binding site.

[0100] In one embodiment, to effectively evaluate the performance of the site prediction model, the random seed for data partitioning is fixed, the validation set is used to adjust the model parameters, the training process is repeated 10 times and the results are saved, and the performance index is calculated using precision, which is as follows:

[0101]

[0102] Where TP stands for true positive, which is the number of correctly identified binding residues among the predicted binding residues; TN stands for true negative, which is the number of correctly identified non-binding residues among the predicted non-binding residues; FP stands for false positive, which is the number of incorrectly identified binding residues among the predicted binding residues; and FN stands for false negative, which is the number of incorrectly identified non-binding residues among the predicted binding residues.

[0103] Furthermore, in order to evaluate the performance of the site prediction model RBSP-M2FN, this example compared RBSP-M2FN with the most advanced site-binding prediction methods (Rsite, Rsite2, RBind, RNAsite and RLBind), and used two independent test sets, TEST9 and TEST18, to evaluate the performance of each method.

[0104] During performance evaluation, since RLBind is the only deep learning method currently available, a random seed was set during model training to ensure that the data in each training set is consistent with that in RBSP-M2FN.

[0105] Table 1 Performance test results of prediction methods on TEST9

[0106]

[0107] Table 2 Performance test results of prediction methods on TEST18

[0108]

[0109] As shown in Tables 1 and 2, RBSP-M2FN achieved the best results in TEST9. On the TEST9 test set, RBSP-M2FN demonstrated exceptional performance, outperforming traditional machine learning methods and surpassing RLBind, another deep learning method. This demonstrates that the multi-dimensional feature fusion network has achieved considerable success in feature fusion, effectively capturing both global and local RNA feature information. In TEST18, it was second only to the best model, RNAsite, with a close difference.

[0110] In addition, if Figure 6 As shown, one embodiment of the present invention further discloses a system for predicting RNA-small molecule binding sites, the system comprising:

[0111] A feature encoding module 10 is configured to obtain global information and local information based on the motifs of the full-length RNA and the motifs of its local neighboring RNAs, wherein the global information includes the sequence and structural feature information of the RNA itself, and the local information includes the sequence and structural feature information of the local neighboring RNAs;

[0112] The prediction module 20 is used to input the global information into a two-dimensional feature fusion network to obtain global features, wherein the two-dimensional feature fusion network includes a convolutional neural network channel and an attention network channel; input the local information into a fully connected neural network to obtain local features; input the fused feature information after integrating the global features and the local features into an Inception module of a one-dimensional feature fusion for deep fusion to obtain deep fusion features; and input the deep fusion features into an output layer to predict the probability value of the binding site.

[0113] This embodiment uses the sequence and structural features of RNA itself as global information, and a limited number of adjacent nucleotide sequences and structural features as local information. These two types of information are fused through a multidimensional feature fusion network, thereby improving the probability of predicting RNA-small molecule binding sites. A two-dimensional feature fusion network is used to perform two-dimensional feature fusion on the global information, and then a one-dimensional feature network is used to perform one-dimensional feature fusion on the local information. These two fusions of features in different dimensions greatly improve the effective utilization of RNA feature information and further reduce the number of parameters, allowing the network to maintain the prediction performance of RNA-small molecule binding sites while also being lightweight.

[0114] In one embodiment, the feature encoding module 10 includes:

[0115] An attribute calculation unit, for calculating input attributes using the motifs of the full-length RNA and the motifs of its local neighboring RNAs;

[0116] An attribute encoding unit, configured to encode the input attributes to obtain a multi-dimensional vector, wherein the vector information includes the sequence-dependent attributes and structure-dependent attributes of the RNA itself and the sequence-dependent attributes and structure-dependent attributes of the local neighboring RNAs;

[0117] A global information calculation unit, configured to obtain the global information based on the sequence-dependent properties and structure-dependent properties of the RNA itself;

[0118] The local information calculation unit is used to obtain the local information based on the sequence dependency properties and structure dependency properties of the local neighbor RNA.

[0119] In one embodiment, the sequence-dependent attributes include nucleotide type and evolutionary conservation score;

[0120] The structure-dependent properties include network topology properties, biochemical properties and ASAs.

[0121] In one embodiment, the convolutional neural network channel includes a two-dimensional convolution layer and a pooling layer, the input of the two-dimensional convolution layer is the global information, and the output of the two-dimensional convolution layer is connected to the pooling layer;

[0122] The attention network channel includes a bidirectional long short-term memory network BiLSTM and an attention network, the input of the BiLSTM is the global information, and the output of the BiLSTM is connected to the attention network;

[0123] The global features are obtained by concatenating the output features of the pooling layer and the output features of the attention network.

[0124] In one embodiment, the BiLSTM includes an input layer, a hidden layer, and an output layer connected in sequence, and the attention network includes a network fusion layer Multiply_1, a softmax layer, and a network fusion layer Multiply_2;

[0125] The hidden layer and output layer of the BiLSTM are both connected to the network fusion layer Multiply_1, the output of the network fusion layer Multiply_1 is connected to the softmax layer, and the output of the softmax layer and the output layer of the BiLSTM are both connected to the network fusion layer Multiply_2.

[0126] In one embodiment, the Inception module includes one-dimensional convolution kernels of different sizes, each of the one-dimensional convolution kernels is arranged in parallel, and each of the one-dimensional convolution kernels is followed by a batch normalization layer.

[0127] In one embodiment, the Inception module includes one-dimensional convolution kernels Conv_1, Conv_2, Conv_3, Conv_4, Conv_5, Conv_6, Conv_8, Conv_8, a pooling layer Pooling_2, and a concatenation layer Concatenate;

[0128] The input of the one-dimensional convolution kernel Conv_1 is the fused feature information, and the output of the one-dimensional convolution kernel Conv_1 is connected to the concatenation layer Concatenate through the batch normalization layer BN_1;

[0129] The input of the one-dimensional convolution kernel Conv_2 is connected to the pooling layer Pooling_2, and the output is connected to the concatenation layer Concatenate through the batch normalization layer BN_2. The input of the pooling layer Pooling_2 is the fused feature information;

[0130] The input of the one-dimensional convolution kernel Conv_3 is the fused feature information, the output of the one-dimensional convolution kernel Conv_3 is connected to the one-dimensional convolution kernel Conv_4 through the batch normalization layer BN_3, and the output of the one-dimensional convolution kernel Conv_4 is connected to the concatenation layer Concatenate through the batch normalization layer BN_4;

[0131] The input of the one-dimensional convolution kernel Conv_5 is the fused feature information. The output of the one-dimensional convolution kernel Conv_5 is connected to the one-dimensional convolution kernel Conv_6 via the batch normalization layer BN_5. The output of the one-dimensional convolution kernel Conv_6 is connected to the one-dimensional convolution kernel Conv_7 via the batch normalization layer BN_6. The output of the one-dimensional convolution kernel Conv_7 is connected to the concatenation layer Concatenate via the batch normalization layer BN_7.

[0132] The output of the concatenation layer Concatenate is connected to the batch normalization layer BN_8 through the one-dimensional convolution kernel Conv_8.

[0133] It should be noted that other embodiments or implementation methods of the RNA-small molecule binding site prediction system of the present invention can refer to the above-mentioned method embodiments, which will not be repeated here.

[0134] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0135] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0136] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for predicting RNA-small molecule binding sites, characterized in that: The method comprises: Based on the motifs of the full-length RNA and the motifs of its local neighbor RNAs, global information and local information are obtained, including: using the motifs of the full-length RNA and the motifs of its local neighbor RNAs to calculate input attributes; encoding the input attributes to obtain a multi-dimensional vector, wherein the vector information includes the sequence dependency attributes and structure dependency attributes of the RNA itself and the sequence dependency attributes and structure dependency attributes of the local neighbor RNAs; based on the sequence dependency attributes and structure dependency attributes of the RNA itself, the global information is obtained; based on the sequence dependency attributes and structure dependency attributes of the local neighbor RNAs, the local information is obtained; wherein the global information includes the sequence and structure feature information of the RNA itself, and the local information includes the sequence and structure feature information of the local neighbor RNAs; Sending the global information into a two-dimensional feature fusion network to obtain global features, wherein the two-dimensional feature fusion network includes a convolutional neural network channel and an attention network channel; Sending the local information into a fully connected neural network to obtain local features; The fused feature information after integrating the global features and the local features is sent to the Inception module of one-dimensional feature fusion for deep fusion to obtain deep fusion features; The deep fusion features are sent to the output layer to predict the probability value of the binding site.

2. The RNA-small molecule binding site prediction method according to claim 1, wherein: The encoding of the input attributes to obtain a multi-dimensional vector includes: The input attributes are encoded from sequence-related attributes, structure-related attributes, network topology attributes and accessible surface areas to obtain a ten-dimensional vector.

3. The RNA-small molecule binding site prediction method according to claim 1, wherein: The sequence-dependent attributes include nucleotide type and evolutionary conservation score; The structure-dependent properties include network topology properties, biochemical properties and ASAs.

4. The RNA-small molecule binding site prediction method according to claim 1, wherein: The convolutional neural network channel includes a two-dimensional convolution layer and a pooling layer, the input of the two-dimensional convolution layer is the global information, and the output of the two-dimensional convolution layer is connected to the pooling layer; The attention network channel includes a bidirectional long short-term memory network BiLSTM and an attention network, the input of the BiLSTM is the global information, and the output of the BiLSTM is connected to the attention network; The global features are obtained by concatenating the output features of the pooling layer and the output features of the attention network.

5. The method for predicting RNA-small molecule binding sites according to claim 4, wherein: The BiLSTM includes an input layer, a hidden layer, and an output layer connected in sequence, and the attention network includes a network fusion layer Multiply_1, a softmax layer, and a network fusion layer Multiply_2; The hidden layer and output layer of the BiLSTM are both connected to the network fusion layer Multiply_1, the output of the network fusion layer Multiply_1 is connected to the softmax layer, and the output of the softmax layer and the output layer of the BiLSTM are both connected to the network fusion layer Multiply_2.

6. The method for predicting RNA-small molecule binding sites according to claim 1, wherein: The step of feeding the local information into a fully connected neural network to obtain local features includes: The local information is sent to the fully connected neural network for dimensionality reduction processing to obtain the local features.

7. The method for predicting RNA-small molecule binding sites according to claim 1, wherein: The Inception module includes one-dimensional convolution kernels of different sizes, each of which is arranged in parallel, and each of which is followed by a batch normalization layer.

8. The method for predicting RNA-small molecule binding sites according to claim 6, wherein: The Inception module includes one-dimensional convolution kernels Conv_1, Conv_2, Conv_3, Conv_4, Conv_5, Conv_6, Conv_8, Conv_8, pooling layer Pooling_2 and concatenation layer Concatenate; The input of the one-dimensional convolution kernel Conv_1 is the fused feature information, and the output of the one-dimensional convolution kernel Conv_1 is connected to the concatenation layer Concatenate through the batch normalization layer BN_1; The input of the one-dimensional convolution kernel Conv_2 is connected to the pooling layer Pooling_2, and the output is connected to the concatenation layer Concatenate through the batch normalization layer BN_2. The input of the pooling layer Pooling_2 is the fused feature information; The input of the one-dimensional convolution kernel Conv_3 is the fused feature information, the output of the one-dimensional convolution kernel Conv_3 is connected to the one-dimensional convolution kernel Conv_4 through the batch normalization layer BN_3, and the output of the one-dimensional convolution kernel Conv_4 is connected to the concatenation layer Concatenate through the batch normalization layer BN_4; The input of the one-dimensional convolution kernel Conv_5 is the fused feature information. The output of the one-dimensional convolution kernel Conv_5 is connected to the one-dimensional convolution kernel Conv_6 via the batch normalization layer BN_5. The output of the one-dimensional convolution kernel Conv_6 is connected to the one-dimensional convolution kernel Conv_7 via the batch normalization layer BN_6. The output of the one-dimensional convolution kernel Conv_7 is connected to the concatenation layer Concatenate via the batch normalization layer BN_7. The output of the concatenation layer Concatenate is connected to the batch normalization layer BN_8 through the one-dimensional convolution kernel Conv_8.

9. A RNA-small molecule binding site prediction system, characterized in that: The system comprises: A feature encoding module is used to obtain global information and local information based on the motifs of the full-length RNA and the motifs of its local neighboring RNAs, wherein the global information includes the sequence and structural feature information of the RNA itself, and the local information includes the sequence and structural feature information of the local neighboring RNAs; A prediction module is configured to feed the global information into a two-dimensional feature fusion network to obtain global features, the two-dimensional feature fusion network including a convolutional neural network channel and an attention network channel; feed the local information into a fully connected neural network to obtain local features; feed the fused feature information after integrating the global features and the local features into an Inception module of a one-dimensional feature fusion for deep fusion to obtain deep fusion features; and feed the deep fusion features into an output layer to predict the probability value of the binding site; The feature encoding module includes: An attribute calculation unit, for calculating input attributes using the motifs of the full-length RNA and the motifs of its local neighboring RNAs; An attribute encoding unit, configured to encode the input attributes to obtain a multi-dimensional vector, wherein the vector information includes the sequence-dependent attributes and structure-dependent attributes of the RNA itself and the sequence-dependent attributes and structure-dependent attributes of the local neighboring RNAs; A global information calculation unit, configured to obtain the global information based on the sequence-dependent properties and structure-dependent properties of the RNA itself; The local information calculation unit is used to obtain the local information based on the sequence dependency properties and structure dependency properties of the local neighbor RNA.

Images