RNA-small molecule binding affinity prediction method based on multi-view learning

Abstract

The invention belongs to the field of intelligent biological prediction, and particularly relates to an RNA-small molecule binding affinity prediction method based on multi-view learning. The method comprises three stages of data set up and down sampling processing, feature construction and extraction based on RNA and small molecule sequences, and binding affinity prediction. According to the method, data set distribution is optimized through an up-down sampling method, multi-view feature information is constructed through sequence information of RNA and small molecules, a deep network is established to extract and depth biological information features, and finally, an MLP regression device is adopted to predict an RNA-small molecule binding affinity value, so that drug discovery work is assisted to be achieved. In order to solve the problem of imbalance of the existing data set, the invention provides a method of combining up and down sampling for processing, and the problem of imbalance is effectively relieved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent biological prediction, specifically involving a method for predicting RNA-small molecule binding affinity based on multi-perspective learning. Background Technology

[0002] Studies have shown that ribonucleic acid (RNA) has become a key player in gene regulation, RNA processing and modification, and catalysis. The functions of RNA molecules are typically achieved through interactions with intracellular components such as proteins, peptides, DNA, and other small molecules. These interactions form the basis for RNA's various biological functions within the cell. RNA has a significant impact on pathology and human health, primarily affecting gene translation, replication, and selective splicing. In past research, utilizing RNA-small molecule interactions to develop targeted drugs has become an important approach for treating such diseases. Compared to traditional protein targets, the structure of RNA makes it a more ideal target for small molecule chemical probes and lead drugs, demonstrating its enormous potential as a therapeutic target for binding small molecules. In summary, in-depth research into the interactions between RNA and small molecules not only opens new pathways for targeted drug development but also provides valuable insights and references for elucidating the mechanisms of RNA-mediated biological processes. Binding affinity refers to the strength or degree of affinity between molecules, typically used to describe the ability of a molecule (such as a drug molecule or ligand) to bind to its target (such as a receptor, enzyme, DNA, RNA, or other biological macromolecules).

[0003] Due to the numerous challenges posed by RNA molecular interactions in biological experimental research, computer-aided design methods have emerged. While RNA-ligand interaction prediction has received widespread attention, work on predicting RNA-small molecule ligand binding affinity is still relatively limited. Recently, the prediction of RNA-small molecule ligand binding affinity has gradually gained interest. Krishnan et al. developed R-SIM, currently the largest RNA-small molecule interaction database, where binding affinity is experimentally calculated. Building upon this, Krishnan et al. developed a computational method based on traditional machine learning, RSAPred, to predict binding affinity using RNA sequence information and molecular descriptors of SMILES. Furthermore, Sun et al. constructed the deep learning framework RLaffinity using 3D structural information of RNA-ligand complexes to predict affinity. While these works have demonstrated effectiveness, their performance in predicting RNA-small molecule binding affinity needs further improvement. Moreover, due to the lack of RNA-small molecule binding affinity data, how to more effectively extract information from limited RNA and SMILES sequences and fully utilize this extracted information for affinity prediction remains a challenge.

[0004] In the field of bioinformatics, a large body of work has been conducted to study the capture of information from RNA sequences and small molecule (SMILES) sequences. For RNA sequences, Su et al. proposed iLoc-lncRNA, which uses 8-mer frequency features to extract features from long non-coding RNA (lncRNA) sequences and uses a binomial distribution to select features for predicting lncRNA subcellular localization. Wang et al. proposed a versatile RNA language modeling framework that combines motif-aware pre-training and type-guided fine-tuning strategies to improve the functional prediction and representation capabilities of RNA sequences. For small molecule (SMILES) sequences, Shao et al. proposed DTIGCCN, which constructs a graph of SMILES sequence information and combines it with protein structural information to predict drug-target interactions. While these methods have demonstrated effectiveness, at the RNA sequence level, the frequency features used do not consider the sequence order information of the RNA sequence, and the language model feature extraction lacks sensitivity to variant sequences and handling of repetitive sequences. For small molecule (SMILES) sequence information, these methods also do not fully consider the information inherent in the topological structure of the atomic arrangement and the physicochemical properties of the sequences themselves. In addition to solving the problem of efficiently extracting RNA sequence information and small molecule SMILES sequence information, the imbalance of RNA subtype category samples is also a problem that needs to be solved for existing RNA-small molecule pair datasets. Summary of the Invention

[0005] This invention proposes EMUDNet, a method for predicting RNA-small molecule binding affinity based on upsampling / downsampling algorithms and multi-view learning. This method can predict the binding affinity of RNA-small molecule interactions, providing effective assistance for RNA-targeted drug development. In this invention, the negative logarithmic form of the dissociation constant is used to standardize the strength of the binding affinity.

[0006] The technical solution of the present invention is as follows: The RNA-small molecule binding affinity prediction method based on multi-view learning includes the following three stages: dataset upsampling and downsampling, feature construction and extraction based on RNA and small molecule sequences, and affinity prediction. The steps of the three stages are as follows: Phase 1: Dataset Upsampling and Downsampling. This phase consists of two steps: upsampling of the minority samples and downsampling of the majority samples. The specific steps are as follows: Step 1: For small sample data An upsampling strategy is adopted, specifically, a combination of adaptive hierarchical clustering and mutation operations is used to supplement the samples to obtain the upsampling results. Specifically, it is expressed as follows:

[0007] In the formula This represents adaptive hierarchical clustering. This indicates a mutation operation. Represent and manipulate.

[0008] Step 2: For multi-sample data By using hierarchical clustering to extract samples, more representative samples are obtained, resulting in downsampling results. Specifically, it is expressed as follows:

[0009] In the formula This represents adaptive hierarchical clustering. Indicates a selection operation. Represent and manipulate.

[0010] The second stage: Feature construction and extraction based on RNA and small molecule sequences. This stage consists of four steps: feature construction and extraction of RNA sequences and feature construction and extraction of small molecule SMILES sequences. The specific steps are as follows: Step 3: First, the RNA sequence is self-trained and embedded using a BERT architecture. Each BERT architecture consists of a self-attention mechanism, a normalization layer, a dropout layer, and a pooling layer, with the input being the RNA sequence. The output is an embedding vector. The specific embedding process is as follows:

[0011] In the formula This indicates the self-attention mechanism. This indicates normalization processing. Indicates the dropout layer. This represents the pooling layer.

[0012] The fourth step is to construct the de Bruijn diagram based on the RNA sequence. Using the GAT model for graph data Feature extraction was performed. The number of heads in the multi-head attention mechanism within the GAT was set to 4, and the average value of the features of all multiple head nodes in the graph was taken as the feature extraction result of the GAT model for the RNA sequence. The mean of all node embeddings was used as the final embedding feature from a graph perspective of the RNA sequence. The calculation process can be represented as follows:

[0013] In the formula, This represents a multi-head graph attention network. This indicates an operation to aggregate and average the values ​​of all nodes. yes The adjacency matrix.

[0014] Step 5: Step 5.1, use Mordred to calculate the molecular descriptor of the small molecule based on the SMILES sequence. Step 5.2 involves setting up an adaptive feature selection module for dimensionality reduction. This module combines a hard sampling strategy with a soft selector and uses the straight-through estimator (STE) technique to ensure the backpropagation process of the model. Step 5.2 consists of three steps: Step 5.2.1 sets up a feature selection module that is compatible with the original descriptor features. Learnable weight matrices of the same dimension Step 5.2.2, will The activation passed to the Sigmoid function makes The weights in the model range from 0 to 1, where the weight data represents the probability of feature selection, thus yielding the soft mapping result. Step 5.2.3 uses a hard binarization strategy to convert the weights to 0 or 1. Specifically, although hard binarization is used during forward propagation, we aim to preserve smooth gradients for effective optimization during backpropagation, thereby ensuring the superiority of feature selection. Therefore, we introduce a pass-through estimator technique to achieve this, namely the detach operation. The calculation process in step 5.2 is as follows:

[0015] in, Represents hard binary operations. This represents the molecular descriptor result matrix after the adaptive feature selector, in the forward propagation, the hard selector Decide which features are retained and which are discarded, especially during backpropagation. Perform a detach operation to allow the gradient to pass through the soft selector. Flow, and thus update As training progressed, It gradually learns which features are most important to the target and optimizes its selection strategy. This gives the model adaptive capabilities, enabling it to optimize feature selection for the task.

[0016] Step 6, which consists of five steps, involves using the rdkit tool to convert the SMILES sequence into an atomic sequence for the small molecule. Step 6.2 uses the topological properties calculated from the atomic diagram as the embedding of each atom, such as tight centrality and betweenness, ultimately yielding the features. Step 6.3: Apply a bidirectional GRU to... Feature extraction is performed to obtain a matrix. Step 6.4 Feature extraction results of GRU Each step involves weighting, i.e., self-attention weighted computation, to obtain the weighted result. Step 6.5 uses mean pooling to obtain a fixed-length output of the entire sequence. That is, take The mean value along the sequence length dimension is used as a feature representation of the entire sequence. The above steps are represented as follows:

[0017] In the formula This represents a bidirectional GRU model. This represents the average pooling layer.

[0018] The third stage: Affinity prediction. Affinity prediction involves two steps: multi-view feature stitching and MLP module prediction.

[0019] The specific steps for this stage are as follows: Step 7: Feature Concatenation. Embed the self-trained BERT into the features. RNA sequence map aggregation features Adaptive selection of molecular descriptor features and atomic-level sequence characteristics of small molecules The components are spliced ​​together to obtain the final RNA-small molecule pair features. .

[0020] Step 8: MLP Prediction. To predict RNA-small molecule binding affinity, this invention constructs a multilayer perceptron structure as the prediction module. Specifically, the prediction module comprises three structures: an input layer, a hidden layer, and an output layer. The input layer receives... Then, weighted calculations are performed through hidden layers. To improve the nonlinearity of the prediction module, the ReLU activation function is used in the hidden layers for transformation operations. Finally, the structure is obtained through the output layer. , which is the negative logarithm of the dissociation constant. The MLP regressor can be expressed as:

[0021] In the formula, It refers to a linear layer. It is the ReLU activation function layer.

[0022] The beneficial effects of this invention are: (1) To address the imbalance problem in existing datasets, this study proposes a combined upsampling and downsampling method to effectively alleviate the imbalance problem.

[0023] (2) Unlike existing methods, this invention uses a multi-view feature extraction method, which to some extent solves the problem of insufficient information extraction in previous methods.

[0024] (3) The superiority and effectiveness of the present invention have been verified on widely used datasets. Attached Figure Description

[0025] Figure 1 This is a framework diagram of the EMUDNet model of the present invention; Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] Figure 1 This invention demonstrates a method for predicting RNA-small molecule binding affinity based on upsampling / downsampling algorithms and multi-view learning. For example... Figure 1 As shown, the workflow of this invention is as follows: First, the dataset is optimized using upsampling and subsampling modules. Then, based on RNA sequences, features are extracted using a self-trained BERT module and a graph feature extraction module. Based on SMILES sequences, depth information of small molecules is extracted using an adaptive molecular descriptor selection module and an atomic-level feature extraction module. Next, the extracted depth features are concatenated and fused. Finally, an MLP regressor is used to predict the depth features.

[0028] Example 1 In Example 1, the present invention constructs and evaluates the model on a public dataset and compares it with the RSApred, DeepRSMA, and EMMPTNet methods. This study aims to compare the prediction error and Pearson coefficient of the models, and the various indicators are shown in Table 1.

[0029] By observing Table 1, we can draw the following conclusions: (1) As shown in Table 1, under the 10-fold cross-validation method, EMUDNet achieved a PCC score of 0.814, higher than EMMPTNet's 0.773; EMUDNet achieved a MAE score of 0.047, which is lower than EMMPTNet's MAE score of 0.053. Therefore, in terms of both linear correlation and accuracy, EMUDNet's average performance is higher than other methods, demonstrating the effectiveness of the model in this chapter. It is worth noting that while the average performance improved, the standard deviation of the EMUDNet model's prediction results increased slightly. This may be due to the use of more complex feature extraction methods such as BERT in the model in this chapter. This may have introduced some noise and perturbation into the model, but for the overall effect, such loss is worthwhile.

[0030] Table 1. Comparison of metrics between EMUDNet and other methods on standard datasets.

[0031] Example 2 To demonstrate the effectiveness of each module of the model, this invention conducts multiple ablation comparison experiments. To verify the effectiveness of the upsampling and downsampling modules, this section will verify it from two aspects: the compatibility experiment of the upsampling and downsampling modules, and the comparison of prediction performance on different RNA subtype data.

[0032] To verify whether the upsampling and downsampling algorithms are effective for the overall model prediction performance, we conducted ablation experiments on the upsampling and downsampling modules based on EMUDNet. The results are shown in Table 2, where EMUDNet indicates that the upsampling and downsampling modules were used, and -UDModule indicates that the upsampling and downsampling modules were not used.

[0033] Table 2 Ablation experiments of upsampling and downsampling modules in EMUDNet

[0034] As shown in Table 2, after implementing the upsampling and downsampling algorithms, EMUDNet shows improvements over -UDModule in both error metrics (RMSE, MAE) and linear correlation (PCC, SPCC), indicating that the upsampling and downsampling modules contribute to improving the model's predictive ability. Furthermore, EMUDNet also demonstrates improved stability in 10-fold cross-validation, suggesting that the upsampling and downsampling modules alleviate overfitting to some extent.

[0035] The main purpose of using the upsampling and downsampling modules is to avoid the model having strong predictive power in a large number of RNA subtype data while ignoring a small number of data, thus leading to overfitting. Therefore, it is necessary to verify the predictive power of each subtype. In this chapter, EMUDNet was used to test the data of 6 different subtypes separately. The results are shown in Table 3. In this table, -UDModule indicates that the upsampling and downsampling modules are not used. Table 3 shows the performance of each subtype data on the MAE index.

[0036] Table 3 Comparison of MAE performance for different data subtypes with and without upsampling / downsampling modules

[0037] MAE is an indicator that measures the absolute error between predicted and true values. As shown in Table 3, after using the upsampling and downsampling modules, the error indicators of the test data on Aptamers, miRNAs, Repeats, Ribosomal RNAs, and Viral RNAs subtypes were all improved, and EMUDNet also showed better performance in terms of standard deviation, i.e., stability. It is worth noting that Aptamers had the lowest error in the -UDModule. After sample balancing, the performance of miRNAs, Repeats, and Viral RNAs outperformed that of Aptamers, demonstrating the effectiveness of the upsampling and downsampling modules. Furthermore, the performance of the Aptamers type did not decrease, indicating that no important information was lost during downsampling, further demonstrating the superiority of combining upsampling and downsampling.

[0038] Then, to verify the effectiveness of the RNA sequence and small molecule features extracted by the EMUDNet method from various perspectives, we combined feature information from different perspectives and used ablation experiments to evaluate the effectiveness of different perspective features. Specifically, we used RNA_Bert to represent the self-trained BERT embedding features of the RNA sequence; GAT_seq to represent the de Bruijn graph features of the RNA sequence through a graph attention network; Ada_Des to represent the adaptive molecular descriptor features of small molecules; and Atomic_Seq to represent the atomic-level sequence features of small molecules. The experimental results are shown in Table 4, where EMUDNet represents the method containing all four perspectives, -RNA_Bert represents the method without the RNA sequence embedding perspective features based on the BERT model, -GAT_seq represents the method without the RNA sequence de Bruijn graph perspective features, -Ada_Des represents the method without the molecular descriptor perspective features of dynamically selected small molecules, and -Atomic_Seq represents the method without the atomic-level feature perspective information of small molecules.

[0039] Table 4 Comparison of experimental results under different combinations of perspective features.

[0040] As shown in Table 4, EMUDNet exhibits the best performance compared to other methods that remove certain perspectives. The features from the GAT_seq and Atomic_Seq perspectives have the greatest impact on prediction performance after ablation. This, to some extent, illustrates the crucial role of sequence information and atomic-level sequence information of SMILES contained in the de Bruijn plot of RNA sequences in predicting RNA-small molecule binding affinity. Furthermore, the PCC and SPCC indices in -GAT_seq and -Atomic_Seq indicate that molecular descriptors and atomic sequence information have a significant impact on the linear correlation of the regression model.

Claims

1. A method for predicting RNA-small molecule binding affinity based on multi-view learning, characterized in that, The steps are as follows: Step 1: For small sample data An upsampling strategy was adopted, using a combination of adaptive hierarchical clustering and mutation operations to supplement the samples, resulting in upsampling results. Specifically, it is expressed as follows: ； In the formula This represents adaptive hierarchical clustering. This indicates a mutation operation. Represent and manipulate; Step 2: For multi-sample data By using hierarchical clustering to extract samples, more representative samples are obtained, resulting in downsampling results. Specifically, it is expressed as follows: ； In the formula This represents adaptive hierarchical clustering. Indicates a selection operation. Represent and manipulate; Step 3: Embed the RNA sequence using a BERT architecture. Each BERT architecture consists of a self-attention mechanism, a normalization layer, a dropout layer, and a pooling layer, with the input being the RNA sequence. The output is an embedding vector. The specific embedding process is as follows: ； In the formula This indicates the self-attention mechanism. This indicates normalization processing. Indicates the dropout layer. Indicates the pooling layer; Step 4: Constructing the de Bruijn diagram based on RNA sequences Using the GAT model for graph data Feature extraction was performed; the number of heads in the multi-head attention mechanism within the GAT was set to 4, and the average value of the features of all multiple head nodes in the graph was taken as the feature extraction result of the GAT model for the RNA sequence. The mean of all node embeddings was used as the final embedding feature from a graph perspective of the RNA sequence. The calculation process can be represented as follows: ； In the formula, This represents a multi-head graph attention network. This indicates an operation to aggregate and average the values ​​of all nodes. yes The adjacency matrix; Step 5: Use Mordred to calculate the molecular descriptor of the small molecule based on the SMILES sequence, and set up an adaptive feature selection module for dimensionality reduction; Step 5.1: Calculate the molecular descriptor of the small molecule using Mordred based on the SMILES sequence. ; Step 5.2: Set up an adaptive feature selection module for dimensionality reduction. This module combines a hard sampling strategy and a soft selector, and uses a pass-through estimator technique to ensure the backpropagation process of the model. The sixth step involves using the rdkit tool to convert the SMILES sequence into an atomic sequence of a small molecule, and using the topological structure properties calculated through the atomic diagram as the embedding of each atom. Then, the GRU and self-attention mechanism are combined to extract features from the temporal information corresponding to the atomic sequence. It is expressed as follows: ； In the formula This represents a bidirectional GRU model. Indicates the average pooling layer; Step 7: Feature stitching; Embedding features with self-trained BERT RNA sequence map aggregation features Adaptive selection of molecular descriptor features and atomic-level sequence characteristics of small molecules The components are spliced ​​together to obtain the final RNA-small molecule pair features. ; Step 8: MLP Prediction Module; A multilayer perceptron structure was constructed as the MLP prediction module, comprising an input layer, a hidden layer, and an output layer. The input layer receives... Then, weighted calculations are performed through hidden layers, where the ReLU activation function is used for transformation operations. Finally, the structure is obtained through the output layer. , which is the negative logarithm of the dissociation constant.

2. The RNA-small molecule binding affinity prediction method based on multi-view learning as described in claim 1, characterized in that, The calculation process in step 5.2 is as follows: ； in, Represents hard binary operations. This represents the molecular descriptor result matrix after the adaptive feature selector; The calculation process for step 5.2 is as follows: ； in, Represents hard binary operations. This represents the molecular descriptor result matrix after the adaptive feature selector, in the forward propagation, the hard selector Decide which features are retained and which are discarded, especially during backpropagation. Perform a detach operation to allow the gradient to pass through the soft selector. Flow, and thus update As training progresses, It gradually learns which features are most important to the target and optimizes the selection strategy; this gives the model the ability to adapt and optimize feature selection in the task.

3. The RNA-small molecule binding affinity prediction method based on multi-view learning as described in claim 1, characterized in that, The specific operation of step 5.2 is as follows: Step 5.2.1 Set a feature similar to the original descriptor Learnable weight matrices of the same dimension ; Step 5.2.2, will The activation passed to the Sigmoid function makes The weights in the model range from 0 to 1, where the weight data represents the probability of feature selection, thus yielding the soft mapping result. ; Step 5.2.3 Use a hard binarization strategy to convert the weights to 0 or 1.

4. The RNA-small molecule binding affinity prediction method based on multi-view learning as described in claim 1, characterized in that, The sixth step, as described above, is performed as follows: Step 6.1 Use the rdkit tool to convert the SMILES sequence into the atomic sequence of the small molecule; Step 6.2: The topological structure properties calculated using the atomic graph are used as the embeddings of each atom, ultimately yielding the features. ; Step 6.3 Apply bidirectional GRU to Feature extraction is performed to obtain a matrix. ; Step 6.4 Feature extraction results of GRU Each step involves weighting, i.e., self-attention weighted computation, to obtain the weighted result. ; Step 6.5 uses mean pooling to obtain a fixed-length output of the entire sequence. That is, take The mean value along the sequence length dimension is used as a feature representation of the entire sequence.

5. The RNA-small molecule binding affinity prediction method based on multi-view learning as described in claim 1, characterized in that, The eighth step, the MLP prediction module, is represented as follows: ； In the formula, It refers to a linear layer. It is the ReLU activation function layer.

Images