A Multi-Task Deep Learning-Based RBP Binding Site Prediction Method

Abstract

This invention discloses an RBP binding site prediction method based on multi-task deep learning, which solves the problem of insufficient training in single-task learning when there are few samples. The key technical solution is to first construct a decoupled dataset suitable for multi-task learning models, then receive and encode RNA sequence samples of various RBP binding sites, establish a multi-task deep learning network framework using a feature sharing network and a task-specific network, and train them alternately. The encoded feature matrix is ​​input into the feature sharing network, and dimensionality reduction is performed using a one-dimensional convolutional neural network algorithm and a long short-term memory algorithm in the feature sharing network. Finally, a multilayer perceptron in the task-specific classification network is used to classify and predict the samples to obtain the classification probability results. This invention's RBP binding site prediction method based on multi-task deep learning can compensate for the lack of samples, improve the robustness of the classification algorithm, and increase the prediction accuracy.

Description

Technical Field

[0001] This invention relates to RNA sequence analysis technology, and in particular to a method for predicting RBP binding sites based on multi-task deep learning. Background Technology

[0002] In recent years, with the development of deep learning, its applications have become increasingly widespread across various fields, including bioinformatics. The improved gene databases have provided ample and comprehensive data for deep learning, giving it unique advantages in bioinformatics and leading to significant progress in various technologies, particularly in the analysis of gene expression and regulation. RNA-binding proteins (RBPs) play a central role in gene expression and regulation analysis. Accurately identifying RBP binding sites on genes is crucial for studying the role of RBPs in gene regulation; therefore, predicting RBP binding sites is an important research objective in bioinformatics.

[0003] In the study of RBP binding sites, validating suitable proteins experimentally within large RBP datasets is crucial. Traditional bioinformatics analysis of RBP binding sites requires extensive biological experiments, placing stringent demands on equipment and making results unpredictable due to gene mutations—a time-consuming and labor-intensive process. Therefore, a technique that reduces time consumption while maintaining accuracy is urgently needed. With the rapid growth of biological data, traditional biological analysis methods face significant challenges. The integration of modern machine learning methods, especially deep learning, promises to greatly improve research efficiency in bioinformatics and potentially uncover hidden patterns within vast amounts of biological data, enabling rapid and accurate predictions.

[0004] Current research on deep learning in RBP binding site prediction has generally achieved considerable results, demonstrating its feasibility in this task. For RBP binding site prediction, the most common deep learning method is to train each RBP individually, i.e., single-task learning. This may yield better results for RBPs with a large number of samples, but is slightly less effective for RBPs with fewer samples, indicating room for improvement. Summary of the Invention

[0005] The purpose of this invention is to provide an RBP binding site prediction method based on multi-task deep learning, which can compensate for the lack of samples, improve the robustness of classification algorithms, and increase prediction accuracy.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0007] A method for predicting RBP binding sites based on multi-task deep learning includes the following steps:

[0008] S1. Construct a decoupled dataset suitable for multi-task learning models;

[0009] S2. Receive RNA sequence samples with multiple RBP binding sites as input, encode the input RNA sequences to obtain the feature matrix of a single sample, and use feature sharing network and task-specific network to establish a multi-task deep learning network framework and perform alternating training.

[0010] S3. Input the encoded feature matrix into the feature sharing network, and use the one-dimensional convolutional neural network algorithm and long short-term memory algorithm in the feature sharing network to reduce the dimensionality and extract the shared sequence feature matrix.

[0011] S4. Use a multilayer perceptron in a task-specific classification network to classify and predict the samples, and obtain the classification probability results.

[0012] In summary, the present invention has the following beneficial effects:

[0013] This invention proposes an RBP-based site prediction method based on multi-task deep learning. Compared to commonly used single-task learning methods, it employs an alternating training approach and introduces a multi-task deep learning algorithm to reduce the impact of insufficient sample size on inadequate training of the classification network, thereby improving the robustness of the classification algorithm. By using a feature-sharing network to share learned network parameter information across multiple similar classification tasks, it can effectively promote mutual learning and improve generalization performance. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the algorithm.

[0015] Figure 2 A schematic diagram showing the number of samples for each type of circular RNA binding protein site;

[0016] Figure 3 A diagram showing the structure of a multi-task deep learning model;

[0017] Figure 4 A plot of AUC results for each RBP on the validation set;

[0018] Figure 5 This is a visual representation of the feature matrix after dimensionality reduction using PCA.

[0019] Figure 6 Line graphs comparing the AUC evaluation results of the three methods. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings.

[0021] According to one or more embodiments, a method for predicting RBP binding sites based on multi-task deep learning is disclosed, such as... Figure 1 As shown, it includes the following steps:

[0022] S1. Construct a decoupled dataset suitable for multi-task learning models;

[0023] S2. Receive RNA sequence samples with multiple RBP binding sites as input, encode the input RNA sequences to obtain the feature matrix of a single sample, establish a multi-task deep learning network framework using feature sharing network and task-specific network, and perform alternating training.

[0024] S3. Input the encoded feature matrix into the feature sharing network, and use the one-dimensional convolutional neural network algorithm and long short-term memory algorithm in the feature sharing network to reduce the dimensionality and extract the shared sequence feature matrix.

[0025] S4. Use a multilayer perceptron in a task-specific classification network to classify and predict the samples, and obtain the classification probability results.

[0026] The RNA-binding protein dataset primarily includes the full-length circular RNA number, the binding site location for each circular RNA, and the start and end positions of each sample within the full-length gene sequence of each circular RNA. The dataset used is a circRNA dataset, comprising 37 types of circular RNAs and approximately 140,000 full-length circRNAs.

[0027] When constructing a positive sample, the starting and ending positions of the binding sites are first calculated, and the midpoint is taken to obtain the center position of the binding sites. Then, the length is extended by 50 bases forward and backward from the midpoint, and finally, an RNA sequence of length 101 is obtained. Since the lengths of each binding site are different, this length can basically cover all binding sites.

[0028] To avoid coupling exclusion issues, circRNA sequences that do not belong to any RBP binding site need to be considered negative samples. When constructing positive samples, each RNA sequence segment is masked at the corresponding positive sample location. After all RBP positive samples have been cut, negative samples are randomly cut from the remaining unmasked circRNA sequences at the same length of 10¹. These negative samples are then evenly distributed among the RBP datasets in a 1:1 ratio, resulting in a final, reasonably decoupled RBP dataset composed of circRNA sequences, as shown below. Figure 2 The figure shows the number of samples for each type of circular RNA binding protein site.

[0029] All input RNA sequences are 101 bases in length. One-hot encoding is used to encode the input RNA sequences, as follows:

[0030] A:[0,0,0,1]

[0031] C:[0,0,1,0]

[0032] G:[0,1,0,0]

[0033] U:[1,0,0,0]

[0034] A stands for adenine, C for cytosine, G for guanine, and U for uracil; these are the four bases that make up the RNA sequence.

[0035] The feature matrix of a single sample is obtained after encoding: R M×N , where M is the length of a single sample sequence, which is 101 here, and N is the dimension of the vector after one-hot encoding, which is 4 here. Therefore, the shape of a single sample matrix input to the neural network is 101×4.

[0036] A multi-task deep learning network framework is built using feature-sharing networks and task-specific networks, such as... Figure 3 As shown, its shared layer network consists of two one-dimensional convolutional neural networks (including pooling layers), one LSTM neural network, and one fully connected network; the task-specific network consists of two fully connected neural networks and an activation function. The activation function is the sigmoid activation function, as shown in the following equation:

[0037]

[0038] The training and test sets were divided in an 8:2 ratio. The Adam optimizer was used to train the model with a learning rate of 1e-4 and the MSE loss function was used as the loss function. The training was conducted for 50 epochs. The MSE loss function is shown in the following formula:

[0039]

[0040] For multiple tasks, namely multiple RBP binding site classification prediction tasks, in the same training, each task will first pass through the feature sharing network, and then pass through the task-specific classification network for that task separately. After the classification prediction result is obtained for each task, the feature sharing network and its corresponding task-specific classification network are updated once.

[0041] The sequence feature matrix, which has already been one-hot encoded, is input into the feature sharing network. First, it undergoes dimensionality reduction through two one-dimensional convolutional layers, embedding the high-dimensional features into the low-level hidden layer space. The first one-dimensional convolutional layer uses a convolutional kernel of length 3 and width 4 to perform a sliding convolution operation along the sequence length, considering the biological characteristic that three bases in an RNA sequence correspond to one codon, and extracting their positional features within local regions.

[0042] The dimensionality of the input single-sample feature matrix is ​​reduced using a one-dimensional convolution algorithm, resulting in the dimensionality-reduced feature matrix: R M’×1 The calculation of M′ is shown in the following formula:

[0043]

[0044] Where M is the length of the input matrix, f is the size of the convolution kernel, p is the width of the padding around the original feature matrix during the convolution operation, and stride is the stride during the convolution operation.

[0045] After two one-dimensional convolutional layers, the input is a bidirectional long short-term memory network with a hidden layer dimension of 200. This fully considers the contextual information of the input RNA sequence along its length, extracting sequence features over a longer range. Hidden features shared by multiple tasks are obtained through a one-dimensional convolutional algorithm and the long short-term memory network.

[0046] When output by the feature sharing network, the dimension of the sample features becomes 400.

[0047] The output obtained from the feature sharing network is used as input and passed through a task-specific classification network for each task. The task-specific classification network consists of a multi-layer perceptron, including two fully connected layers, and finally the classification probability for the task is obtained.

[0048] A multilayer perceptron consists of the following components:

[0049] The first layer consists of a fully connected layer with 400-dimensional input and 128-dimensional output, using the ReLU activation function;

[0050] The second layer consists of a fully connected layer with 128-dimensional input and 1-dimensional output, using the sigmoid activation function.

[0051] This invention uses AUC as the evaluation metric. AUC is defined as the area under the ROC curve and the coordinate axis, typically ranging from 0.5 to 1. A UC closer to 1.0 indicates higher validity, while a UC closer to 0.5 indicates lower validity. The ROC curve, or Receiver Operating Characteristic curve, is plotted using a binary classification method (here, the cutoff value), with the true positive rate (the probability of classifying a true positive as positive) on the ordinate and the false positive rate (the probability of classifying a true negative as positive) on the abscissa.

[0052] The experiment was conducted, and the results are as follows: Figure 4 As shown, AUC represents the model's AUC value on the test set. Experimental results demonstrate that this invention performs well in classification prediction on most RBPs, especially those with fewer samples, achieving an average AUC of 0.9383 on the test set. This result significantly surpasses the findings of previous methods.

[0053] To further investigate the rationality of the method in this invention, for each RBP, the feature matrix of the last layer of the shared network is extracted and reduced to two dimensions using PCA. This matrix is ​​then mapped to a Cartesian coordinate system for visualization analysis. Four RBPs are randomly selected for comparison, such as... Figure 5 As shown in the figure, graph analysis reveals concentrated clusters and clear classification boundaries, with only a very small number of samples being misclassified, consistent with the AUC evaluation results for each RBP.

[0054] To demonstrate the superiority of the method in this invention, it is compared with other methods in the field of RBP binding site prediction, with AUC as the comparison metric. The method used for comparison here, CRIP, comes from the paper "CRIP: predicting circRNA–RBP-bindingsites using a codon-based encoding and hybrid deep neural networks". CRIP is a single-task learning method, characterized by its use of codon-based one-hot encoding: in an RNA sequence, every three bases correspond to a codon, resulting in 64 possible codons (3 stop codons and 61 codons corresponding to 20 amino acids). These 20 amino acids and the stop codons are used as 21-dimensional features of the feature vector for one-hot encoding. For an input RNA sequence of length 101 nt (base units), a sliding window of length 3 is used, with each window yielding one set of codons, resulting in 99 sets of codons after the sliding window is completed. The amino acid list corresponding to each codon is then one-hot encoded in 21 dimensions, resulting in an input feature matrix of shape 99×21 for a single sample pair. Three methods were compared here: ①CMTL: the multi-task deep learning method proposed in this invention, where the encoding method is 4-dimensional one-hot encoding; ②CRIP: a single-task deep learning method, where the encoding method is 21-dimensional one-hot encoding based on codons; ③CECMTL: using the multi-task deep learning model proposed in this invention, where the encoding method is 21-dimensional one-hot encoding based on codons. The AUC results of the three methods in RBP binding site prediction were compared, as follows... Figure 6 As shown, CECMTL and CMTL achieve better prediction accuracy than CRIP on most RBPs, especially on RBPs with fewer samples, such as TAF15 and QKI. They are only weaker than CRIP on RBPs with more samples, such as AGO2 and FMRP. This indicates that multi-task learning methods are significantly better than single-task learning methods when the number of samples is insufficient. This improvement mainly comes from their feature sharing ability between similar tasks.

[0055] This multi-task deep learning-based RBP combined with site prediction method employs an alternating training approach compared to commonly used single-task learning methods. It also proposes a multi-task deep learning algorithm to mitigate the impact of insufficient sample size on inadequate training of the classification network, thereby improving the robustness of the classification algorithm. Furthermore, a feature-sharing network is used to share learned network parameters across multiple similar classification tasks, effectively promoting mutual learning and enhancing generalization performance.

[0056] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A method for predicting RBP binding sites based on multi-task deep learning, characterized in that, It includes the following steps: S1. Construct a decoupled dataset suitable for multi-task learning models; S2. Receive input RNA sequence samples with multiple RBP binding sites, encode the input RNA sequences to obtain the feature matrix of a single sample, and establish a multi-task deep learning network framework using a feature-sharing network and a task-specific network, and perform alternating training. Step S2 specifically includes: The input RNA sequence is encoded using one-hot encoding to obtain the feature matrix R of a single sample. M×N Where M is the length of a single sample sequence, i.e., the number of bases it contains, and N is the dimension of the vector after one-hot encoding, specifically encoded as A:[0,0,0,1]. C:[0,0,1,0] G:[0,1,0,0] U:[1,0,0,0] A stands for adenine, C for cytosine, G for guanine, and U for uracil; these are the four bases that make up the RNA sequence. A multi-task deep learning network framework is established using feature-sharing networks and task-specific networks; In the same training session, for multiple RBP binding site classification prediction tasks, each task first passes through a feature sharing network and then through a task-specific classification network for that task. After each task obtains a classification prediction result, the feature sharing network and its corresponding task-specific classification network are updated once. S3. Input the encoded feature matrix into the feature sharing network, and use the one-dimensional convolutional neural network algorithm and long short-term memory algorithm in the feature sharing network to reduce the dimensionality and extract the shared sequence feature matrix. S4. Use a multilayer perceptron in a task-specific classification network to classify and predict the samples, and obtain the classification probability results.

2. The RBP binding site prediction method based on multi-task deep learning according to claim 1, characterized in that, The decoupled dataset in step S1 is specifically constructed as follows: S11. Construct positive samples. On the circular RNA sequence corresponding to the original dataset, take the midpoint between the start and end positions of the binding site, and then extend the length by 50 bases forward and backward from the midpoint to obtain an RNA sequence of length 101. S12. For each segment of RNA sequence, cover the corresponding position on the full-length circular RNA; S13. After cutting all the positive samples of RBP, cut the negative samples on the remaining unmasked circular RNA sequences at the same length of 101, and distribute them to the RBP dataset in a 1:1 ratio of positive and negative samples to obtain a decoupled dataset composed of circular RNA sequences.

3. The RBP binding site prediction method based on multi-task deep learning according to claim 1, characterized in that, Step S3 is as follows: S31. Use a one-dimensional convolution algorithm to reduce the dimensionality of the input single sample feature matrix. After two one-dimensional convolutional layers, the high-dimensional features are embedded into the low-dimensional hidden layer space to obtain the dimensionality-reduced feature matrix: R M’×1 The formula for calculating M' is shown below: Where M is the length of the input matrix, f is the size of the convolution kernel, p is the width of the padding around the original feature matrix in the convolution operation, and stride is the stride in the convolution operation. S32. Use a long short-term memory network to further reduce the dimensionality of the feature matrix to obtain the dimensionality-reduced feature matrix.

4. The RBP binding site prediction method based on multi-task deep learning according to claim 3, characterized in that, Step S4 specifically involves: The task-specific classification network consists of a multi-layer perceptron, including two fully connected neural networks, which output classification probabilities; A multilayer perceptron consists of the following components: The first layer consists of a fully connected layer with 400-dimensional input and 128-dimensional output, using the ReLU activation function; The second layer consists of a fully connected layer with 128-dimensional input and 1-dimensional output, using the sigmoid activation function.

Images