Construction method for predicting RNA and small molecule binding site model based on RNA sequence and structure information characteristics and application of construction method

Abstract

The invention discloses a construction method and application of a model for predicting binding sites of RNA and small molecules based on RNA sequence and structure information characteristics, and belongs to the technical field of biology. The construction method of the model comprises the following steps: S1, collecting modeling data; s2, dividing a data set; s3, dividing positive and negative data sets; s4, combining the RNA characteristics; and S5, constructing a model. The invention further provides a small molecule binding site prediction method based on RNA sequence and structure information. The prediction method is realized by adopting the model constructed by the method. The method has the beneficial effect that the efficiency and the accuracy of predicting the RNA-small molecule binding site are greatly improved. Experimental results on two independent test sets show that the overall performance of RNABind is superior to that of an existing method, and particularly, the call is greatly improved, so that the model can better capture a real sample. RNABind is expected to become a useful tool for designing RNA targeted small molecule drugs.

Description

Technical Field

[0001] The present invention relates to the field of biotechnology, and in particular to a method for constructing a model for predicting RNA-small molecule binding sites based on RNA sequence and structural information features and an application thereof. Background Art

[0002] In recent years, RNA small molecules have gradually attracted attention as potential targets for treating diseases. RNA-small molecule binding sites are crucial to the research of targeted drugs. However, due to the high complexity of RNA sequences and structures, predicting RNA-small molecule binding sites remains a huge challenge.

[0003] Because existing methods do not have a high degree of recognition for real samples, we proposed a new prediction method based on the sequence and structure characteristics of RNA: RNABind. Experimental results on two independent test sets show that RNABind's overall performance is better than existing methods, especially greatly improving recall, allowing the model to better capture real samples. RNABind is expected to become a useful tool for designing RNA-targeted small molecule drugs.

[0004] The functions of RNA molecules are complex and diverse, and they play a vital role in maintaining life activities and body functions. RNA small molecules can interact with specific binding sites of RNA macromolecules to regulate gene expression, interfere with RNA structure, and other functions. In the field of drug research and development, RNA-targeted small molecules have specific binding capabilities to RNA structure or function, can be used to regulate the biological activity of RNA, and are expected to be used as new drugs for the treatment of related diseases. This has promoted the development of computational methods for predicting RNA-small molecule binding sites.

[0005] Due to the diversity of RNA sequences and structures, and the scarcity of RNA-small molecule binding site data, only a few methods have been developed, including Rsite, Rsite2, RBind, RNAsite, and RLBind. Rsite is a computational method based on RNA tertiary structure. It first calculates the Euclidean distance between each nucleotide and all other nucleotides in the RNA molecule, and then determines the extreme point in the distance curve as the nucleotide of the functional site. Rsite2 uses the secondary structure of RNA for prediction based on Rsite. RBind is a structure-based computational network that uses degree and binding proximity to identify RNA-small molecule binding sites. RNAsite and RLBind both use RNA sequence and structural features to predict RNA-small molecule binding sites. As for the two latest methods, although they both use the sequence and structural features of RNA, the recognition of real samples is not very high. Summary of the invention

[0006] The technical problem to be solved by the present invention is how to provide a method for constructing a model for predicting the binding sites of RNA and small molecules.

[0007] The present invention solves the above technical problems through the following technical means:

[0008] The first aspect of the present invention provides a method for constructing a model for predicting RNA-small molecule binding sites based on RNA sequence and structural information features, comprising the following steps:

[0009] S1. Modeling data collection: The IDs of RNA-small molecule interacting RNA complexes were collected from the PDB database and used as modeling samples after processing;

[0010] S2. Dataset division: All RNA-small molecule interaction samples are divided into training set and test set according to the proportion;

[0011] S3. Division of data into positive and negative sets: the atomic distance between nucleotides and small molecules is less than a specific distance cutoff value Then the nucleotides are defined as those that interact with the small molecule, i.e., positive samples, and the rest are negative samples;

[0012] S4. Combination of RNA features: Select the one-hot coding of RNA sequence and the evolutionary conservation score of each nucleotide in the sequence as sequence features; select site prior properties, network topology properties, biochemical properties and ASA coding as structural features; combine the features to obtain comprehensive features with RNA sequence and structural information;

[0013] S5. Model construction: A deep learning model consisting of a convolution module, a sliding window module, and a self-attention module is used to predict the comprehensive features with RNA sequence and structural characteristics and obtain results.

[0014] Preferably, in S1, the processing method is that if the RNA molecule forms a complex with one or more ligands, the complex is selected for binding site prediction.

[0015] Preferably, in S1, the processing method is to use structural similarity clustering to remove redundant structures, and the cutoff value of pairwise structural similarity TM-scoreRNA is 0.3.

[0016] Preferably, in S1, the processing method is to simulate the tertiary structure of RNA and retain the RNA by filtering RNA structures with multi-strand or pseudoknot interactions.

[0017] Preferably, in S5, a convolution module is used to extract the global features of RNA encoding; a sliding window module is used to extract the local features of RNA encoding; and finally, the global and local features are spliced ​​and input into the self-attention module to obtain the prediction result.

[0018] The second aspect of the present invention proposes a method for predicting small molecule binding sites based on RNA sequence and structure information, and the prediction method is implemented using the model constructed by the above method.

[0019] Preferably, the method comprises the following steps:

[0020] (1) Computationally encode each nucleotide sequence and structural feature of RNA;

[0021] (2) splicing each nucleotide sequence of the RNA with the structural features and inputting them into the small molecule binding site prediction model to obtain the probability value of each nucleotide of the RNA being a small molecule binding site;

[0022] (3) Determine whether the nucleotide is a small molecule binding site based on a threshold value, and determine it as a binding site if it is greater than the threshold value.

[0023] Preferably, in step (1), the sequence features include one-hot encoding of the sequence, and using the ConSurfDB algorithm to determine the evolutionary conservation score of each nucleotide; the structural features include using a secondary structure similarity site prior algorithm to calculate the probability of the binding site in the RNA secondary structure, using a tertiary structure network topology algorithm to calculate the importance of each node in the tertiary structure, and using the RNAsol algorithm to calculate the ASA of the nucleotide.

[0024] Preferably, in step (2), the encoding length of each RNA nucleotide is 16D, and the fixed length of each RNA is 64, which is input into the model to obtain 64×16 feature information for each RNA.

[0025] Preferably, in step (3), 64×16 feature information of each RNA is input into the model to determine whether the model is a small molecule binding site for each nucleotide.

[0026] The beneficial effects of the present invention are:

[0027] The present invention proposes a new prediction method RNABind to predict RNA-small molecule binding sites. In this work, we use the sequence, secondary structure, tertiary structure and biochemical properties of RNA. The characteristics of the sequence include one-hot encoding of nucleotide types and evolutionary conservation scoring of nucleotides. The characteristics of the structure include the site prior properties of secondary structure similarity, the network topological properties of the tertiary structure and the accessible surface area (ASA), among which the secondary structure is an encoding method designed by us to extract features from RNAs with similar secondary structures. Finally, the biochemical properties include the encoding of RNA molecular weight and side chain Pka. The efficiency and accuracy of predicting RNA-small molecule binding sites are greatly improved. The experimental results on two independent test sets show that the overall performance of RNABind is better than the existing methods, especially the recall is greatly improved, so that the model can better capture the real samples. RNABind is expected to become a useful tool for designing RNA-targeted small molecule drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a schematic diagram of the construction process of the model for predicting RNA and small molecule binding sites in Example 1 of the present invention;

[0029] Figure 2 This is a 3D diagram of the prediction results in Example 1 of the present invention. Figure 2 A is a 3D graph of the prediction results from one perspective. Figure 2 B is a 3D graph of the prediction results from another perspective;

[0030] Figure 3 This is a comparison chart of the prediction method RNABind of the present invention and other existing prediction methods in Example 1 of the present invention, wherein Figure 3 a is the comparison chart of Precision prediction index, Figure 3 b is the comparison chart of Recall prediction index, Figure 3 c is the comparison chart of AUC prediction index, Figure 3 d is the comparison chart of MCC prediction indicators. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] Unless otherwise specified, the test materials and reagents used in the following examples can be obtained from commercial sources.

[0033] If no specific techniques or conditions are specified in the examples, they can be carried out according to the techniques or conditions described in the literature in the art or according to the product instructions.

[0034] Embodiment 1:

[0035] A method for constructing a model for predicting RNA-small molecule binding sites based on RNA sequence and structural information features, comprising the following steps:

[0036] S1. Modeling data collection:

[0037] Extract the complex data of RNA and small molecule interactions from the PDB (Protein Data Bank) database, including the collection of three-dimensional structural information containing RNA and small molecule ligands, as well as related RNA-small molecule interaction data.

[0038] Through the preprocessing steps: data cleaning and format conversion, the data is organized into a format suitable for modeling. To ensure the integrity and accuracy of the data, incomplete or poor quality data is removed. The processed data is used as a modeling sample.

[0039] S2. Dataset division:

[0040] The RNA-small molecule interaction sample data set obtained in step S1 is divided into a training set, a validation set, and a test set in a ratio of 70%, 15%, and 15%. (That is, 70% of the data is used for training, 15% for validation, and 15% for testing.)

[0041] The diversity and representativeness of the data were taken into consideration during the division process to ensure that each data set comprehensively covers the different characteristics and patterns of RNA-small molecule binding.

[0042] S3. Division of positive and negative data sets:

[0043] According to the set atomic distance cutoff Determine whether the nucleotides in RNA interact with the atoms of the small molecule. If the atomic distance between the nucleotide and the small molecule is less than Then the nucleotide is marked as a positive sample; otherwise, it is marked as a negative sample.

[0044] For negative samples, it is also necessary to ensure that they represent the non-interactive part of RNA binding to small molecules in order to balance the ratio of positive and negative samples in the data set.

[0045] S4. Combination of RNA features:

[0046] Sequence features: Select one-hot encoding of RNA sequences to represent the specific sequence information of each nucleotide; calculate the evolutionary conservation score of each nucleotide in the sequence to describe the degree of conservation of the nucleotide.

[0047] Structural features: Extract site prior properties (such as labeling information of known binding sites), network topological properties (such as the connection of nucleotides), biochemical properties (such as chemical environment), and ASA (accessible surface area) encoding, which are used to describe the structural characteristics of RNA.

[0048] The above features are combined to form comprehensive features with RNA sequence and structure information for subsequent model training and prediction.

[0049] S5. Model construction:

[0050] The prediction model is constructed using deep learning methods, and the model architecture includes convolution module, sliding window module and self-attention module.

[0051] Convolutional module: used to extract local patterns and important features in RNA sequence and structural features.

[0052] Sliding window module: Processes sequence data through a sliding window mechanism to capture local information under different windows.

[0053] Self-attention module: used to capture global dependencies and important information in sequence and structural features to improve prediction accuracy.

[0054] When training the model, the comprehensive features are input into the deep learning model for training, and the prediction accuracy is improved by adjusting the model parameters and optimization algorithms (such as Adam optimizer).

[0055] After the model training is completed, the test set is used to evaluate the model and calculate the prediction performance indicators (such as accuracy, recall, F1 score) to verify the effectiveness and generalization ability of the model.

[0056] Based on the test results and performance indicators, the model is further optimized, including adjusting model parameters, adding or reducing features, improving network structure, etc.

[0057] The optimized model is applied to the actual task of predicting the binding sites of RNA and small molecules, helping to provide effective prediction tools in fields such as scientific research and drug development.

[0058] The model construction process diagram is as follows Figure 1 As shown, from Figure 1 As can be seen from the figure, RNA is input into the model after being encoded in sequence and structure, and the results are spliced ​​after being processed by the convolution module and the sliding window module respectively, and finally input into the self-attention module to obtain the model prediction result.

[0059] A method for predicting small molecule binding sites based on RNA sequence and structure information is implemented using the model constructed by the above method.

[0060] The specific steps include:

[0061] (1) Calculate and encode each nucleotide sequence and structural feature of RNA respectively; sequence features include one-hot encoding of the sequence, and use of the ConSurfDB algorithm to determine the evolutionary conservation score of each nucleotide; structural features include using the secondary structure similarity site prior algorithm to calculate the probability of binding sites in the RNA secondary structure, using the tertiary structure network topology algorithm to calculate the importance of each node in the tertiary structure, and using the RNAsol algorithm to calculate the ASA of the nucleotide;

[0062] (2) Each nucleotide sequence of the RNA is spliced ​​with the structural features and input into the small molecule binding site prediction model to obtain the probability value of each RNA nucleotide being a small molecule binding site; the encoding length of each RNA nucleotide is 16D, and the fixed length of each RNA is 64. After input into the model, each RNA contains 64×16 feature information.

[0063] (3) The 64×16 feature information of each RNA is input into the model, and the threshold is used to determine whether the nucleotide is a small molecule binding site. If the value is greater than the threshold, it is determined to be a binding site, thereby deriving whether the model is a small molecule binding site for each nucleotide.

[0064] In order to evaluate the performance of the model, we use four indicators: precision (pre), recall (recall), AUC and Matthews correlation coefficient (MCC) to comprehensively evaluate the overall performance of the model. To reduce the error of model training, we repeatedly trained the model 10 times and then selected the model with the best indicators for analysis.

[0065] pecision=(|TP|) / (|TP|+|FP|)

[0066]

[0067] Where TP and TN represent true positives and true negatives, respectively, which here refer to the number of binding residues and non-binding residues correctly predicted by the model, and FP and FN represent false positives and false negatives, respectively, which here refer to the number of incorrectly predicted binding residues and non-binding residues. AUC is defined as the area under the ROC curve, which is calculated by plotting the false positive rate (FPR) and the true positive rate (TPR). MCC is widely used to measure the performance of binary classification models, especially in the case of imbalanced data. Here we use it to evaluate the comprehensive classification ability of the model.

[0068] Result analysis:

[0069] In the embodiments of this article, the sequence and structural features of the RNA we used are complementary to a certain extent, which provides more comprehensive information for predicting the binding sites of RNA-small molecules. In order to verify this idea, we conducted ablation experiments on structural and sequence features. As shown in Table 1 below, after removing the site prior properties of the secondary structure or the network topological properties of the tertiary structure, the prediction effect has deteriorated overall, especially the site prior properties of the secondary structure, which greatly improves the recall index and enables the model to better identify real samples. In addition, from a biological point of view, the secondary structure and tertiary structure of RNA contain information that sequence features do not have, such as pairing relationships between bases, spatial structures, etc. This information can provide a more comprehensive feature description and can complement the sequence information, thereby improving the performance of model prediction.

[0070]

[0071]

[0072] Table 1: Feature ablation table on test set T18

[0073] Importance of prior properties of secondary structure similarity sites

[0074] In order to prove the importance of the prior properties of the secondary structure similarity points, we statistically analyzed the different running results after adding and removing this feature, and then compared the results of each RNA in the T18 test set, and performed statistical comparisons. We compared the precision and recall indicators of each RNA with the latest existing RLBind calculation method. We found that the Precision and Recall of RNABind in the T18 test set are mostly higher than RLBind. In addition, we found that the Precision and Recall of 4pqv and 2mis predicted in RLBind were 0.111, 0.083 and 0.4, 0.182 respectively, while in RNABind the Precision and Recall reached 0.429, 0.75 and 0.875, 0.636 respectively, which further verified the importance of the prior properties of the secondary structure similarity points.

[0075] Comparison with existing methods

[0076] In order to evaluate the superiority of our method, we compared RNABind with the existing state-of-the-art methods (Rsite2, RBind and RNAsite). Two test sets, Test18 (a test set containing 18 RNAs) and Test9 (a test set containing 9 RNAs), were used to evaluate the performance of different computational methods. We used four indicators, Precision, Recall, AUC and MCC, for comparison. Precision and Recall are two indicators that are often referenced together in the evaluation of classification models, while MCC is more comprehensive in evaluating the classification results of imbalanced data. Figure 3 As shown, we can see that the comprehensive performance of RNABind is better than other calculation methods, especially in the indicators of Recall and MCC. On the T18 test set, RNABind's Recall is 0.751, which is much higher than Rsite2 (0.214), RBind (0.173) and RNAsite (0.263), and RNABind MCC is 0.557, which is also better than Rsite2 (0.010), RBind (0.187) and RNAsite (0.253). In addition, although the Precision indicator is lower than RNAsite on the T9 test set, it is better than other calculation methods in the other three indicators. In short, after we fully utilize the sequence and structural features of RNA, the comprehensive performance of RNABind is better than the existing calculation methods.

[0077] Models Precision Recall AUC MCC Rsite2 0.370 0.214 0.010 0.474 RBind 0.655 0.173 0.187 0.559 RNAsite 0.675 0.263 0.253 0.776 RNABind 0.683 0.751 0.828 0.554

[0078] Table 2: T18 test set

[0079] Models Precision Recall AUC MCC Rsite2 0.382 0.187 0.072 0.528 RBind 0.681 0.230 0.278 0.592 RNAsite 0.750 0.378 0.426 0.790 RNABind 0.683 0.806 0.871 0.517

[0080] Table 3: T9 test set

[0081] Further research

[0082] The identification of binding sites between RNA and small molecules facilitates the design of targeted RNA drugs, for example, metal ions represent key elements of RNA structure and function. In the Neurospora VS ribozyme (PDB code: 2mis), metal ions play multiple roles; they are important for substrate recognition, formation of the active site, and altering the pKa of two key nucleobases that contribute to the general acid-base mechanism. Under the prediction of RNABind, 7 of the 11 binding site nucleotides of RNA2mis in the test set T18 were correctly predicted with a recall of 0.636. Guanidino-neomycin B recognizes the frameshift site stem-loop in the HIV-1 RNA helix (PDB code: 2juk), which is an attractive target for drug development. Surprisingly, RNABind was able to correctly predict 12 of the 13 binding sites in 2juk with a recall of 0.923. (e.g. Figure 2 As shown, the red part is the part with correctly predicted binding sites, and the blue part is the part with incorrectly predicted binding sites)

[0083] in conclusion:

[0084] In this paper, a new method for predicting small molecule RNA binding sites that combines sequence and structure information, RNABind, is proposed. By extracting the sequence and structural features of RNA molecules, the sequence-based features include one-hot encoding of RNA nucleotide types and evolutionary conservation scores for each nucleotide, and the sequence-based features include secondary structure similarity point prior properties, network topological properties of tertiary structures, releasable surface areas (ASAs) and biochemical properties. We use a neural network based on convolution and self-attention. The network is divided into two branches, one is a convolution module to extract the global features of RNA, and the other is a sliding window module to extract the neighboring features of RNA. They are then simply spliced ​​and input into the self-attention module, and finally output after a linear layer. After experiments, we found that RNABind has a comprehensive performance that is much better than other methods in two independent test sets T18 and T9. This may be because we reasonably use the sequence and structural features of RNA and use a superior model. In fact, identifying small molecule binding sites in RNA can promote the development of targeted drug design. In particular, models with more predicted true positives and fewer predicted false positives will speed up the process, and our model can better identify true samples and better help researchers identify the binding sites of RNA and small molecules.

[0085] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a model for predicting RNA-small molecule binding sites based on RNA sequence and structural information features, characterized in that: The following steps are involved: S1. Modeling data collection: The IDs of RNA-small molecule interacting RNA complexes were collected from the PDB database and used as modeling samples after processing; S2. Dataset division: All RNA-small molecule interaction samples are divided into training set and test set according to the proportion; S3. Division of data into positive and negative sets: the atomic distance between nucleotides and small molecules is less than a specific distance cutoff value Then the nucleotides are defined as those that interact with the small molecule, i.e., positive samples, and the rest are negative samples; S4. Combination of RNA features: One-hot coding of RNA sequences and the evolutionary conservation score of each nucleotide in the sequence were selected as sequence features; site prior properties, network topology properties, biochemical properties and ASA coding were selected as structural features; After combining the features, comprehensive features with RNA sequence and structure information are obtained; S5. Model construction: A deep learning model consisting of a convolution module, a sliding window module, and a self-attention module is used to predict the comprehensive features with RNA sequence and structural characteristics and obtain results.

2. The construction method according to claim 1, characterized in that: In S1, the processing method is that if the RNA molecule forms a complex with one or more ligands, the complex is selected for binding site prediction.

3. The construction method according to claim 1, characterized in that: In S1, the processing method is to use structural similarity clustering to remove redundant structures, and the cutoff value of pairwise structural similarity TM-scoreRNA is 0.

3.

4. The construction method according to claim 1, characterized in that: In S1, the processing method is to simulate the tertiary structure of RNA and retain RNA by filtering RNA structures with multi-strand or pseudoknot interactions.

5. The construction method according to claim 1, characterized in that: In S5, a convolution module is used to extract the global features of RNA encoding; a sliding window module is used to extract the local features of RNA encoding; and finally, the global and local features are spliced ​​and input into the self-attention module to obtain the prediction result.

6. A method for predicting small molecule binding sites based on RNA sequence and structure information, characterized in that: The prediction method is implemented by using a model constructed by the method described in any one of claims 1 to 5.

7. The prediction method according to claim 6, characterized in that: The following steps are involved: (1) Computationally encode each nucleotide sequence and structural feature of RNA; (2) splicing each nucleotide sequence of the RNA with the structural features and inputting them into the small molecule binding site prediction model to obtain the probability value of each nucleotide of the RNA being a small molecule binding site; (3) Determine whether the nucleotide is a small molecule binding site based on a threshold value, and determine it as a binding site if it is greater than the threshold value.

8. The prediction method according to claim 1, characterized in that: In the step (1), the sequence features include one-hot encoding of the sequence, and using the ConSurfDB algorithm to determine the evolutionary conservation score of each nucleotide; The structural features include using the secondary structure similarity site prior algorithm to calculate the probability of the binding site in the RNA secondary structure, using the tertiary structure network topology algorithm to calculate the importance of each node in the tertiary structure, and using the RNAsol algorithm to calculate the ASA of the nucleotide.

9. The prediction method according to claim 6, characterized in that: In step (2), the encoding length of each RNA nucleotide is 16D, and the fixed length of each RNA is 64. When input into the model, each RNA contains 64×16 feature information.

10. The prediction method according to claim 6, characterized in that: In the step (3), 64×16 feature information of each RNA is input into the model to determine whether each nucleotide is a small molecule binding site.

Images