Definer: A computational method for accurate identification of RNA pseudouridine sites based on deep learning

doi:10.21203/rs.3.rs-3103674/v1

Download PDF

Research Article

Definer: A computational method for accurate identification of RNA pseudouridine sites based on deep learning

https://doi.org/10.21203/rs.3.rs-3103674/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

RNA molecules are important biomolecules in cells and play a crucial role in processes such as genetic information transfer and gene expression regulation. However, in addition to their basic nucleotide composition, RNA molecules undergo various modifications, including pseudouridine. Pseudouridine is a critical site of alteration that is found in many non-coding RNAs and has a role in a number of biological processes, including gene expression, RNA structural stability, and the development of several illnesses. Accurate identification of pseudouridine sites in RNA molecules is of significant importance for understanding their functionality and regulatory mechanisms. Traditional experimental methods often rely on techniques such as chemical modifications and mass spectrometry analysis. However, these methods are costly, time-consuming, and limited in terms of sample size. Therefore, the development of an efficient and accurate computational method for identifying pseudouridine sites in RNA holds great scientific significance and practical application value.

Results

In this study, we propose a deep learning-based computational method, Definer, to accurately identify RNA pseudouridine loci in three species, H. sapiens, S. cerevisiae and M. musculus. The method incorporates two sequence coding schemes, including NCP and One-hot, and then feeds the extracted RNA sequence features into a deep learning model constructed from CNN, GRU and Attention. The benchmark dataset contained data from three species, namely H. sapiens, S. cerevisiae and M. musculus, and the results using 10-fold cross-validation showed that the model accuracy reached 82.95, 86.01 and 87.15 for the three species, respectively, with Definer significantly outperforming other existing methods. Meanwhile, the data sets of two species, H. sapiens and S. cerevisiae, were tested independently to further demonstrate the predictive ability of the model.

Conclusion

It is well known that RNA modifications are an important component of gene regulation and most biological processes depend on RNA modifications. Among them, pseudouridine modification is one of the crucial modification sites, and the accurate identification of pseudouridine sites in RNA is important for understanding their functions and regulatory mechanisms. Therefore, this paper proposes a new predictor, Definer, which can accurately identify pseudouridine sites in three species: H. sapiens, S. cerevisiae, and M. musculus. results on benchmark and independent test sets show that Definer has good performance over other existing methods and can accurately identify pseudouridine sites in cross-species data sets. set can accurately identify pseudouridine loci. In addition, we have developed software that runs on a local computer to provide users with a better presentation of the pseudouridine site prediction process and results for the three species datasets.

RNA pseudouridine sites

feature extraction

deep learning

sequence analysis

RNA modifications are a key component of gene regulation and participate in many biological processes[1, 2]. So far, more than 150 types of RNA modifications have been discovered in the field of biology [3, 4]. One of the most frequent RNA modifications is the pseudouridine () modification, which was first identified in the 1960s. It can be found in a number of non-coding RNAs, such as mRNA, tRNAs, and snRNA, among others[5, 6]. The most common processes of RNA modification are pseudouridylation and methylation [7]. Studies have shown that pseudouridine can change the secondary and tertiary structure of RNA, affect the speed of gene expression, and is closely related to various diseases, such as Parkinson's disease, congenital keratinization disorder, and myelodysplastic syndrome keratosis, which are associated with pseudouridine modification mutations[8, 9]. Therefore, the investigation of pseudouridine modification sites holds great significance for both biology and medicine[10, 11]. With the advent of the post-genomic era, there has been a rapid increase in genomic data, and traditional biological experiments are no longer sufficient to meet actual research needs[12]. Therefore, more convenient computational models need to be developed to accurately extract and identify pseudouridine modification site information[13, 14].

For predicting pseudouridine sites in H. sapiens, S. cerevisiae, and M. musculus, numerous computational techniques based on machine learning and deep learning have been developed. Li et al. [15] constructed the first prediction model for pseudouridine sites, PPUS, based on the SVM algorithm for predicting Homo sapiens and Saccharomyces cerevisiae through a web server. Chen et al. [16] constructed datasets for H. sapiens, S. cerevisiae, and Mus musculus and combined the PseDNC encoding method with SVM to build the iRNA-PseU prediction model. The constructed dataset was further used in subsequent research. He et al. [17] proposed PseUI, which combines five encoding methods with SVM algorithm and further improves the accuracy of pseudouridine site recognition by applying sequence forward feature selection. Subsequently, Tahir et al. [18] developed iPseU-CNN, a computational model that utilizes One-hot encoding and two-layer convolutional neural networks to predict pseudouridine modification sites. Liu et al. [19] proposed a novel sequence-based method XG-PseU for predicting pseudouridine modification sites in RNA. The method employs the Extreme Gradient Boosting (XGBoost) algorithm to encode the sequences and establishes a classification model to identify the pseudouridine sites. Bi et al. [20] discovered the ensemble learning algorithm, which integrates five different machine learning classifiers to build the EnsemPseU predictor for predicting pseudouridine sites. In order to improve prediction performance, A new predictor known as RF-PseU was developed by Lv et al. [21] to optimize the feature representation and obtain a better model, an optical gradient lifter algorithm and an incremental feature selection strategy are used to select the best feature space vector for training the random forest model RF-PseU. Mu et al. [22] constructed MU-PseUDeep by combining the original sequence and secondary structure with a convolutional neural network, further improving the performance of predicting pseudouridine sites. Song et al. [23, 24] constructed PIANO and PSI-MOUSE predictors based on genomic and sequence features for predicting pseudouridine sites. This is the first time that genomic-derived features have been introduced and achieved good performance in predicting pseudouridine sites. Li et al. [25] developed the porpoise predictor based on the stacked ensemble learning method and used four feature selection methods.

There is still a great deal of opportunity for improvement in forecasting RNA sequence performance based on current approaches, despite the major contributions and foundational work offered by earlier studies[26, 27]. Developing better prediction methods will enable a comprehensive understanding of the relationship between RNA sequences and life activities. The majority of currently available prediction tools use a single feature extraction approach and conventional machine learning algorithms[28, 29]. Due to the extremely complex sequence features exhibited by biological sequences, traditional machine learning methods cannot achieve better prediction performance. Deep learning algorithms have strong learning and generalization abilities, and possess good modeling capabilities [30]. Therefore, it is necessary to improve prediction performance by increasing sequence features and developing more suitable classification algorithms[31].

To address the issues mentioned above, this paper introduces Definer, a deep learning-based computational method for accurately identifying pseudouridine (Ψ) sites in three different species: H. sapiens, S. cerevisiae, and M. musculus. Firstly, we combined the One-hot and NCP feature encoding schemes to extract RNA sequence information. Secondly, using three deep learning models—the convolutional neural network (CNN), the gated recurrent unit (GRU), and the attention mechanism—we built site prediction models, Definer. Finally, ten-fold cross-validation and independent testing showed that, compared with state-of-the-art methods, Definer significantly improved the prediction performance on Ψ site identification in all three species.

2.1 The entire experimental framework

Figure 1 illustrates the five primary steps that make up the experimental design of this work: data collecting, feature extraction, model construction, performance evaluation, and development of visualization software[32]. Firstly, benchmark datasets and independent test sets for three species, H. sapiens, S. cerevisiae, and M. musculus, were collected from relevant literature and public databases [24]. Secondly, to extract sequence information from the datasets, two feature extraction methods were utilized. Subsequently, a deep learning-based predictor, Definer, was constructed, which achieved good performance on all three species. Furthermore, we evaluated and compared Definer with several existing methods and observed a significant improvement in its prediction performance. Finally, we developed a visualization software and made it publicly available for online use, which facilitates data processing and analysis for users.

2.2 Data collection

In order to facilitate comparison with existing methods, we used the dataset constructed by Chen et al [16], which is commonly used in most of the existing prediction methods, such as iRNA-PseU [16], PseUI [17], iPseU-CNN [18], XG-PseU [19], and Porpoise [25]. The datasets for the three species were obtained from the RMBase database [33], including three benchmark datasets H. sapiens (H_990), S. cerevisiae (S_628), and M. musculus (M_994), which were used for model training, and two independent test sets, which only included H. sapiens (H_200) and S. cerevisiae (S_200) species. Table 1 displays the datasets' specifics. Both the dataset and the model code are accessible at https://github.com/louise1223/Definer.

Table 1

Datasets introduction
Datasets	Positive sample count	Negative sample count	Sequence length
H_990	495	495	21
S_628	314	314	31
M_994	497	497	21
H_220	100	100	21
S_220	100	100	31

2.3 Feature Extraction

Feature extraction is a critical stage in the creation of a prediction model. Its objective is to digitize feature vectors from RNA sequence fragments for processing and analysis. During the feature extraction process, it is necessary to encode the four nucleotides Adenine (A), Cytosine (C), Guanine (G), and Uracil (U) into digitized feature vectors, so that they can be applied to the prediction model[34]. The way of extracting input data has a great impact on the model. Only by choosing a suitable feature extraction method according to specific conditions can better training results be achieved. Efficient feature extraction methods can effectively extract more representative feature vectors and provide strong support for subsequent model construction[35]. Two feature extraction techniques were used in this study: one-hot encoding and nucleotide chemical characteristics. These two feature extraction methods are briefly explained below.

2.3.1 One-hot Encoding

One-hot is a binary encoding method, and is also one of the basic feature representation methods for RNA sequences. Converting categorical variables into numerical data that models can process is a common use. This encoding method converts each possible value of a categorical variable into a binary array[36] .

For example, let is say we have a car's color attribute, with possible values of "red", "blue", and "green". Using One-hot encoding, we can convert this attribute into a binary array of three elements. For instance, if the car's color is "blue", the encoded result would be [0,1,0], where the second element is 1, indicating that the car is blue.

In some machine learning models, using raw categorical variables may cause problems such as in linear regression and logistic regression. One-hot encoding can solve this problem by converting categorical variables into equivalent numeric data, allowing the model to process these variables correctly. It is important to note that when there are many possible values for the categorical variable, using One-hot encoding can lead to a sharp increase in data dimensions, which may affect the efficiency and accuracy of the model. Therefore, in practical applications, it is necessary to decide whether to use One-hot encoding and how to encode it based on specific circumstances.

The basic idea of using One-hot encoding for feature extraction of a sequence is to convert each nucleotide of the sequence into a four-dimensional binary vector, where only one position in the vector is 1 and the rest are 0. The four nucleotides A, U, C, and G will be converted into vectors (1,0,0,0), (0,1,0,0), (0,0,1,0), and (0,0,0,1), respectively.

2.3.2 NCP

The nucleotide chemical properties (NCP) encoding method is a method of converting RNA sequence information into feature vectors, mainly used for sequence classification and prediction problems. This encoding method numerically represents the chemical properties of each nucleotide in the RNA sequence and converts them into a fixed-length feature vector[37]. Specifically, the NCP encoding method represents the chemical properties of each nucleotide as a triplet (x, y, z), where x represents the nucleotide's charge property, y represents its hydrophobicity property, and z represents its hydrogen bonding property. Then, normalize each triplet of nucleotides to a range between 0 and 1, and use it as an element of the feature vector[38]. For an RNA sequence of length L, arrange the feature vectors composed of triplets of each nucleotide to obtain an L x 3 matrix, which is the feature matrix after NCP encoding. Based on the three different chemical structural properties, nucleotides in RNA sequences can be represented by a vector ${L}_{j}=({X}_{j},{Y}_{j},{Z}_{j})$, where X represents cyclic structure, Y represents hydrogen bonding, and Z represents chemical functionality. The feature representation method of NCP is shown in Eq. (1).

$$\begin{array}{c}\left\{\begin{array}{c}{X}_{j}=\left\{\begin{array}{c}1,if{L}_{j}ϵ\left(A,G\right)\\ 0,if{L}_{j}ϵ\left(C,U\right)\end{array}\right.\\ {Y}_{j}=\left\{\begin{array}{c}1,if{L}_{j}ϵ\left(A,C\right)\\ 0,if{L}_{j}ϵ\left(G,U\right)\end{array}\right.\\ {Z}_{j}=\left\{\begin{array}{c}1,if{L}_{j}ϵ\left(A,U\right)\\ 0,if{L}_{j}ϵ\left(C,G\right)\end{array}\right.\end{array}\right.\#\left(1\right)\end{array}$$

The NCP encoding technique describes the chemical characteristics of each nucleotide in the RNA sequence more precisely than One-hot encoding, which can better capture the differences in sequence information. At the same time, the NCP encoding method also avoids the problem of high dimensionality in One-hot encoding. The NCP encoding method has been applied to various RNA sequence analysis and prediction tasks and has practical value.

2.4 Classifier

2.4.1 Overall framework of the model(Definer)

Convolutional neural network (CNN), gated recurrent unit (GRU), and attention are three popular deep learning models that will be employed in this study to create a prediction model for RNA pseudouridine sites. Firstly, the input data is processed through the first convolution layer of the CNN, which performs cross-correlation operations on the matrix of each channel from left to right and top to bottom using convolution kernels. Then, the obtained data is regularized to prevent overfitting, as shown in equations (2) and (3).

$$\begin{array}{c}Conv（\text{x}）=\sum _{p=1}^{P}\sum _{q=1}^{Q}{w}_{pq}^{k}{x}_{i}+p,q\#\left(2\right)\end{array}$$

$$\begin{array}{c}L={E}_{in}+\lambda \sum _{j}\left|{W}_{j}\right|\#\left(3\right)\end{array}$$

Next, the data and parameters are compressed through the pooling layer of CNN, and the compressed data is fed into the second layer of CNN. At the same time, the feature tensor is fused with GRU and the Relu activation function is used to accelerate the convergence of the model, as shown in Eq. (4). The update gate ${Z}_{t}$is used to filter information, and ${W}_{z}$ controls the retention level of new and old information input at each time step. The reset gate ${r}_{t}$ is used to filter information, and ${W}_{r}$ controls the retention level of input information at each position at time t-1 [39].

$$\begin{array}{c}\left\{\begin{array}{c}{Z}_{t}=\sigma \left({W}_{z}\bullet \left[{h}_{t-1},{x}_{t}\right]\right)\\ {r}_{t}=\sigma \left({W}_{r}\bullet \left[{h}_{t-1},{x}_{t}\right]\right)\\ {\stackrel{\sim}{h}}_{t}=\text{tanh}\left(W\bullet \left[{r}_{t}*{h}_{t-1},{x}_{t}\right]\right)\\ {h}_{t}=\left(1-{Z}_{t}\right)*{h}_{t-1}+{z}_{t}*{\stackrel{\sim}{h}}_{t}\end{array}\right.\#\left(4\right)\end{array}$$

Then, following the same method of combining the second layer convolution of CNN with GRU, the feature tensors of the third layer convolution of CNN and GRU are fused. Finally, to focus on important information and fully absorb it, an Attention mechanism is added to the model.

2.4.2 CNN

Convolutional neural network (CNN) has also been extensively applied in areas including natural language processing, genetic sequence analysis, and drug discovery since the emergence of deep learning techniques. A feedforward neural network known as a CNN may analyze two- or three-dimensional input, including pictures and genetic sequences.

The following paragraph explains the typical input and operations of a convolutional neural network (CNN). The input to a CNN is usually a two-dimensional or three-dimensional tensor, such as a three-dimensional tensor for RGB images or a two-dimensional tensor for gene sequences. In the convolutional layer, the network uses a set of learnable filters to perform convolution operations on the input data, in order to extract local features. The filters are usually small two-dimensional or three-dimensional tensors. The convolution operation can be viewed as applying the filter to all positions of the input data, resulting in a feature map. The convolutional layer usually includes multiple filters to extract different features. After the convolutional layer, a pooling layer is typically added to reduce the size of the feature map, in order to reduce computation and memory requirements. The pooling operation usually performs subsampling on each feature map, for example, the max pooling operation divides the feature map into non-overlapping rectangular regions and takes the maximum value in each region as the downsampled value. The network then converts the feature maps generated by the convolutional and pooling layers into a vector in the fully connected layer, where it employs one or more fully connected layers to map the feature maps to the output. Each neuron in the completely linked layer is often connected to every other neuron in the layer above.

When using CNNs, it is typically necessary to preprocess the input data, such as encoding gene sequences. Then, the CNN can be trained using training data and evaluated using testing data. The gradient of the network weights is often calculated during training using the backpropagation algorithm, and the weights are updated using an optimization approach.

2.4.3 Gate Recurrent Unit

The following paragraph explains the Gated Recurrent Unit (GRU), a commonly used variant of recurrent neural networks (RNNs) that effectively addresses issues such as vanishing and exploding gradients when processing sequential data. Compared to traditional RNNs, GRU introduces reset gates and update gates to control information propagation and retention, thereby better capturing long-term dependencies in sequential data. The computation of the reset and update gates is shown in Eq. (5).

$$\begin{array}{c}\left\{\begin{array}{c}{r}_{t}=\sigma \left({x}_{t}{W}_{xr}+{H}_{t-1}{W}_{hr}+{b}_{r}\right)\\ {Z}_{t}=\sigma \left({x}_{t}{W}_{xz}+{H}_{t-1}{W}_{hz}+{b}_{z}\right)\end{array}\right.\#\left(5\right)\end{array}$$

where $\sigma$ is the sigmoid function, ${r}_{t}$ denotes the output of the reset gate, and ${z}_{t}$ denotes the output of the update gate.

To enhance the model's performance even further, GRU can be stacked to create a more complex neural network. It is often necessary to encode the input into a GRU network-friendly format when employing a GRU network. Specifically, input can be encoded using word vectors, character vectors, and other methods, and then fed into the GRU network. Additionally, hyperparameters of the GRU network such as network depth, hidden layer size, learning rate, regularization, etc., need to be set.

2.4.4 Attention Mechanism

Attention is a common neural network model that mainly functions to weight and sum the input data, allowing the model to focus more on the parts of the data that are useful for the current task. Attention was first applied to image classification problems in computer vision and later widely used in fields such as natural language processing, speech recognition, and genetic sequences. The schematic diagram of the Attention structure is shown in Fig. 2.

In this structure, each element of the input sequence has a corresponding value vector, while the query vector is the key information used to calculate the weights. At each time step, Attention calculates the weight of each input element, and then combines them by weighted summation to obtain the output vector. This process can be implemented in the following three steps: first, calculate the similarity between each query vector and each value vector of the input elements to obtain a weight vector; next, calculate the weight by normalizing the weight vector of each input element with respect to each query vector, obtaining a weight distribution for weighted summation; finally, perform the weighted summation, multiply each value vector of the input elements by the corresponding weight and sum all results by weighted summation to obtain the output vector.

By stacking several attention layers, deeper neural networks can be created, further enhancing the model's performance. Hyperparameters like the number of attention heads, size of hidden layers, learning rate, regularization, etc. must also be set when using attention mechanisms.

2.5 Performance Evaluations

When establishing a predictive model, it is crucial to evaluate and validate its feasibility. Model evaluation often employs three methods: k-fold cross-validation, independent testing, and overlapping testing. These methods can comprehensively and accurately test the model on different datasets, to obtain reliable performance evaluation and conclusions. Therefore, choosing an appropriate evaluation method is very important when establishing a predictive model. To facilitate comparison with existing methods and highlight the effectiveness of the proposed method, we selected the first two methods for evaluation. Specifically, we employed 10-fold cross-validation and independent tests on the training and testing datasets, respective. For a classification task, accuracy (ACC) is the most basic evaluation index, which represents the percentage of correct classification. However, the basic evaluation index often cannot reflect the model's performance well, which may lead to poor judgments. This study uses four evaluation indicators, including specificity (Sp), sensitivity (Sn), accuracy (ACC), and Matthew’s correlation coefficient (MCC) to evaluate the predictive model [40, 41]. The calculation formulas for the four evaluation indicators are shown below.

$$\begin{array}{c}{S}_{n} =\frac{TP}{TP+FN}\#\left(6\right)\end{array}$$

$$\begin{array}{c}{S}_{p}=\frac{TN}{TN+FP}\#\left(7\right)\end{array}$$

$$\begin{array}{c}ACC=\frac{TP+TN}{TP+TN+FP+FN}\#\left(8\right)\end{array}$$

$$\begin{array}{c}MCC=\frac{TP\times TN-FP\times FN}{\sqrt{\left(TP+FP\right)\left(TP+FN\right)\left(TN+FP\right)\left(TN+FN\right)}}\#\left(9\right)\end{array}$$

In this paragraph, TP, FP, TN, and FN respectively represent the number of true positives, false positives, true negatives, and false negatives.

3.1 Distribution of nucleotide positions at pseudouridine sites

To analyze the characteristics of pseudouridine sites in RNA sequences, we used Two Sample Logo [42] to calculate the importance of nucleotides at each position. Two Sample Logo is a tool for computing differences between nucleotide samples and the significance of nucleotides at each position in a sequence. The nucleotide distributions of pseudouridine sites in the H. sapiens, S. cerevisiae, and M. musculus species are shown in Figs. 3(A), 3(B), and 3(C), respectively. The size of each letter represents the frequency of the corresponding base at that position, with larger letters indicating higher frequencies. At each position, the letters are arranged in order of dominance from top to bottom, with the most dominant base at the top. From the figures, it can be seen that in H. sapiens, uridine (U) content is highest near the central position 11, while cytidine (C) is mainly distributed at downstream positions 17 and 20. In S. cerevisiae, guanine (G) is mostly distributed in the upper-middle region, while uridine (U) is distributed at the central positions 14, 15, and 16. Adenine (A) is mainly distributed in the upper-middle region, with three consecutive A bases at positions 13, 14, and 15. In M. musculus, uridine (U) is distributed in the upper-middle region, with U bases at positions 9, 11, 12, and 13. These results indicate that there are different nucleotide distribution patterns between pseudouridine and non-pseudouridine sites in the H. sapiens, S. cerevisiae, and M. musculus species, and therefore, it is necessary to establish a universal prediction model across different species.

3.2 Performance comparison analysis of different feature extraction methods

An efficient feature extraction method can effectively extract more representative feature vectors and provide strong support for subsequent model construction [43]. This section compared One-hot, NCP, and their fusion, respectively, by placing these three feature extraction methods into the predictor for comparison. Tables 2, 3, and 4, respectively, display the comparative findings for the three species H. sapiens, S. cerevisiae, and M. musculus.

Table 2

Three feature extraction methods used in H_990 comparison of results
Feature extraction	ACC(%)	Sn(%)	Sp(%)	MCC(%)
One-hot	81.02	83.25	78.78	62.38
NCP	76.55	76.77	76.33	53.48
One-hot + NCP	82.95	85.28	80.62	66.43
Note: Bold numbers indicate the maximum value in this column.

Table 3

Three feature extraction methods used in S_628 comparison of results
Feature extraction	ACC(%)	Sn(%)	Sp(%)	MCC(%)
One-hot	85.81	85.00	86.62	72.18
NCP	84.52	86.28	82.77	69.91
One-hot + NCP	86.01	84.05	87.97	72.66
Note: Bold numbers indicate the maximum value in this column.

Table 4

Three feature extraction methods used in M_944 comparison of results
Feature extraction	ACC(%)	Sn(%)	Sp(%)	MCC(%)
One-hot	86.93	88.32	85.54	74.03
NCP	85.76	85.97	85.56	71.76
One-hot + NCP	87.15	87.04	87.26	74.41
Note: Bold numbers indicate the maximum value in this column.

From the table, it can be observed that both encoding methods combined perform better on all three datasets than using a single feature extraction method. On the H_990 dataset, the fusion of One-hot and NCP outperforms both single feature extraction methods in all four evaluation metrics, with an accuracy of 82.95%, which is 1.93% higher than that of One-hot alone. Although the single feature extraction method Sn outperforms the fusion on the S_628 and M_944 datasets, the fusion performs much better than the single feature extraction method on the other three evaluation metrics. The fused ACCs on the S_628 and M_944 datasets are 86.01% and 87.15%, respectively. Therefore, we choose the fusion of One-hot and NCP to extract sequence feature information for pseudouridine sites.

3.3 Performance comparison analysis of different models

The classifier plays a significant role in the experiment and is closely tied to the outcome of the experiment. Experiment performance can be significantly enhanced by developing a proper predictive model. Deep learning is a machine learning algorithm with feature learning ability. It can extract and learn low-level data features to obtain more abstract high-level features. In recent years, the genomics databases have grown rapidly. Only by using classifiers with stronger learning ability can we better learn and mine effective information in huge databases. In this section, a predictor called Definer was constructed based on commonly used deep learning algorithms. We compared it with several traditional machine learning algorithms and commonly used deep learning algorithms, including SVM, RF, LightGBM, and CNN, based on ten-fold cross-validation on three benchmark datasets. The comparison results of the three benchmark datasets are shown in Fig. 4. It can be seen that the evaluation indicators of deep learning algorithms are much higher than those of traditional machine learning algorithms. We built a new predictor Definer based on CNN, which integrates two GRUs and introduces Attention. From the prediction results, our model is superior to other classifiers in all four evaluation indicators.

3.4 Comparative analysis on Independent Datasets

We carried out additional validation and evaluation on two independent test sets, consisting of H. sapiens and S. cerevisiae, in order to further demonstrate the efficacy of our proposed technique, Definer, in predicting pseudouridine sites. We additionally evaluated the efficacy of Definer in relation to numerous extant techniques on the same independent test sets, and the outcomes were succinctly outlined in Table 5. Notably, our predictor achieved significantly better performance than other methods in all four evaluation metrics on both test datasets. Particularly, in H. sapiens, our method showed remarkable accuracy (ACC) reaching 83.50%, which was 6% higher than the best performing existing method, Porpoise. These results further attest to the robustness and reliability of Definer in predicting pseudouridine sites across different species.

Table 5

Comparison Evaluation with state-of-the-art several methods on the same independent test dataset
Species	Method	ACC(%)	Sn(%)	Sp(%)	MCC(%)
H.sapiens	XG-PseU[19]	67.50	68.00	67.00	35.00
(H_200)	iPseU-CNN[18]	69.00	77.72	60.81	40.00
	PseUI [17]	65.50	63.00	68.00	31.00
	iRNA-PseU [12]	61.50	58.00	65.00	23.00
	EnsemPseU [20]	69.50	73.00	66.00	39.00
	RF-PseU [21]	75.00	78.00	72.00	50.00
	Porpoise [25]	77.35	82.30	72.40	55.13
	Definer	83.50	91.00	76.00	73.34
S.cerevisiae	XG-PseU [19]	71.00	75.00	67.00	42.14
(S_200)	iPseU-CNN [18]	73.50	68.76	77.82	47.00
	PseUI [17]	68.50	65.00	72.00	37.00
	iRNA-PseU [12]	60.00	63.00	57.00	20.00
	EnsemPseU [20]	75.00	85.00	65.00	51.00
	RF-PseU [21]	77.00	75.00	79.00	54.00
	Porpoise [25]	83.50	88.00	79.00	67.27
	Definer	86.00	90.00	82.00	73.45
Note: Bold numbers indicate the maximum value in this column.

3.5 Comparison with other existing methods

This section presents a comparison between Definer and the latest methods on three benchmark datasets, including XG-PseU [19], iPseU-CNN [18], PseUI [17], iRNA-PseU [16], EnsemPseU [20], RF-PseU [21], and Porpoise [25], and the comparison results are shown in Table 6. As can be seen from the table, when performing ten-fold cross-validation on the same two independent test datasets, Definer outperforms the other 7 prediction methods on all four evaluation metrics in two species, S. cerevisiae and M. musculus, and in H. sapiens, the Sn of the porpoise tool is higher than our predictor, but our predictor is far superior to porpoise in the other three evaluation metrics, indicating that Definer is more accurate and stable than other methods. Combining the comparison results on the independent test set in the previous section, it can be concluded that Definer can accurately predict pseudouridine sites in the three species, H. sapiens, S. cerevisiae, and M. musculus.

Table 6

Comparison of three benchmark datasets with existing methods
Species	Method	ACC(%)	Sn(%)	Sp(%)	MCC(%)
H.sapiens	XG-PseU	65.44	63.64	67.24	31.00
(H_990)	iPseU-CNN	66.68	65.00	68.78	34.00
	PseUI	64.24	64.85	63.64	28.00
	iRNA-PseU	60.40	61.01	59.80	21.00
	EnsemPseU	66.28	63.46	69.09	33.00
	RF-PseU	64.30	66.10	62.60	29.00
	Porpoise	78.53	89.11	67.94	58.45
	Definer(10-fold CV)	82.95	85.28	80.62	66.43
S.cerevisiae	XG-PseU	68.15	66.84	69.45	37.00
(S_628)	iPseU-CNN	68.15	66.36	70.45	37.00
	PseUI	65.13	62.74	67.52	30.00
	iRNA-PseU	64.49	64.65	64.33	29.00
	EnsemPseU	74.16	73.88	74.45	49.00
	RF-PseU	74.80	77.20	72.40	49.00
	Porpoise	81.69	81.21	82.17	63.38
	Definer(10-fold CV)	86.01	84.05	87.97	72.66
M.musculus	XG-PseU	72.03	76.48	67.57	45.00
(M_944)	iPseU-CNN	71.81	74.79	69.11	44.00
	PseUI	70.44	74.58	66.31	41.00
	iRNA-PseU	69.07	73.31	64.83	38.00
	EnsemPseU	73.85	75.43	72.25	48.00
	RF-PseU	74.80	73.10	76.50	50.00
	Porpoise	77.75	77.83	77.67	55.55
	Definer(10-fold CV)	87.15	87.04	87.26	74.41
Note: Bold numbers indicate the maximum value in this column and CV indicates cross-validation.

3.6 Software engineering

Software engineering and web servers are popular in the Internet age. In order to facilitate users to identify potential pseudo-uridine sites from RNA sequences, a software visualization based on our model Definer has been developed in this section [40]. The software is developed based on the Python Tkinter framework and the main interface of the software is shown in Fig. 5.

On the main interface, users can enter or copy the RNA sequence they want to query into the input box in the lower right corner, or upload a txt file. Clicking the "example" button allows users to view the required format for the txt file. After a file is successfully uploaded, the data content will be displayed in the lower right box. The software provides three models for users to choose from, named after the corresponding species: H. sapiens, S. cerevisiae, and M. musculus. When users click the "Confire" button above the text box, the model will analyze and calculate the sequence, and return the sequence's name, length, and whether it contains a site to the user. After the prediction is complete, users can click the "download" button to export the prediction result file to a specified path. The prediction result interface is shown in Fig. 6. The software is freely accessible at https://github.com/louise1223/Definer.

Numerous biological processes are known to depend heavily on RNA modifications. The discovery of pseudouridine alteration sites is one of these modifications that is crucial since it can shed light on the underlying mechanisms. Such insights hold significant value for both the fields of biology and medicine. To this end, our study presents a novel predictor, Definer, which accurately identifies pseudouridine sites across three species: H. sapiens, S. cerevisiae, and M. musculus. Definer combines two encoding schemes, including NCP and One-hot, and then builds a deep learning model based on the combination of CNN, GRU, and Attention. The model integrates CNN and GRU to fuse feature tensors and adds Attention to fully focus on key information and learn. Finally, the features are classified, and the model is evaluated through ten-fold cross-validation and independent testing. Benchmark datasets and independent test outcomes demonstrate that our predictor has good performance, surpassing other existing methods and accurately identifying pseudouridine sites in cross-species datasets. In addition, we have developed software that runs on a local computer, providing users with a better display of the pseudouridine site prediction process and results for the three species datasets. Of course, there is still room for improvement for Definer. With the rapid increase in genomic data, collecting RNA pseudouridine sites for new species, building new datasets, and adding encoding schemes to extract more effective information from RNA sequences can all improve the accuracy of RNA modification site prediction.

Ethics approval and consent to participate

Not applicable.

Consent to publication

All authors give consent to publish.

Availability of data and materials

The datasets for the three species, the Definer model code, and the Software engineering code are available for free download on the https://github.com/louise1223/Definer.

Competing interests

The authors declare that we have no known competing financial interests or personal relationships that could influence the work reported in this article.

Funding

This work is supported by the Natural Science Foundation of Jilin Province (YDZJ202301ZYTS401,YDZJ202301ZYTS288),the Science and Technology Project of the Education Department of Jilin Province (JJKH20220245KJ, JJKH20220226SK), the National Natural Science Foundation of China Joint Fund Project (U19A200496).

Authors’ contributions

Xin Feng: Conceived the study, designed and performed the experiment, and wrote the original draft. Ping Wang: Provided input on method design, evaluation, case studies, and the manuscript. Zhansen Ma: Provided input in method design and evaluation. Jiezhang Wu: Provided input in the software and evaluation sections. Ruihao Xin: Conceived the study and edited the manuscript.

Acknowledgments

We thank all individuals who participated in this study.

Jack K, Bellodi C, Landry DM, Niederer RO, Dinman JD: rRNA pseudouridylation defects affect ribosomal ligand binding and translational fidelity from yeast to human cells. Molecular Cell 2011, 44(4):660-666.
Carlile TM, Rojas-Duran MF, Zinshteyn B, Shin H, Bartoli KM, Gilbert WV: Pseudouridine profiling reveals regulated mRNA pseudouridylation in yeast and human cells. Nature 2014, 515(7525):143-146.
Pelchat M: Binding site of Escherichia coli RNA polymerase to an RNA promoter. Biochemical and Biophysical Research Communications 2004.
Maroney PA, Romfo CM, Nilsen TW: Nuclease protection of RNAs containing site-specific labels: a rapid method for mapping RNA-protein interactions. RNA 2001, 6(12):1905-1909.
Basak A, Query C: A pseudouridine residue in the spliceosome core is part of the filamentous growth program in yeast. Cell Reports 2014, 8(4).
Mei YP, Liao JP, Shen J, Yu L, Liu BL, Liu L, Li RY, Ji L, Dorsey SG, Jiang ZR: Small nucleolar RNA 42 acts as an oncogene in lung tumorigenesis. Oncogene 2012, 31(22):2794-2804.
Cohn WE: Pseudouridine, a Carbon-Carbon Linked Ribonucleoside in Ribonucleic Acids: Isolation, Structure, and Chemical Characteristics. Journal of Biological Chemistry 1960, 235:1488.
Li X, Ma S, Yi C: Pseudouridine: the fifth RNA nucleotide with renewed interests. Current Opinion in Chemical Biology 2016, 33:108-116.
Chan CM, Huang RH: Enzymatic characterization and mutational studies of TruD – the fifth family of pseudouridine synthases. Archives of Biochemistry & Biophysics 2009, 489(1-2):15-19.
Karijolich J, Yu YT: Converting nonsense codons into sense codons by targeted pseudouridylation. Nature.
Zhang Y, Hamada M: DeepM6ASeq: Prediction and characterization of m6A-containing sequences using deep learning. BMC Bioinformatics 2018, 19(S19).
Quinlan AR, Hall IM: Quinlan AR, Hall IM. BEDTools: a flexible suite of utilities for comparing genomic features. Bioinformatics 2010, 26(6):841-842.
Swami A, Jain R: Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 2013, 12(10):2825-2830.
Godzik LA: Cd-hit: a fast program for clustering and comparing large sets of protein or nucleotide sequences. Bioinformatics 2006, 22(13):1658.
Li Y-H, Zhang G, Cui Q: PPUS: a web server to predict PUS-specific pseudouridine sites. Bioinformatics 2015, 31(20):3362-3364.
Chen W, Tang H, Ye J, Lin H, Chou K-C: iRNA-PseU: Identifying RNA pseudouridine sites. Molecular Therapy-Nucleic Acids 2016, 5:e332.
He J, Fang T, Zhang Z, Huang B, Zhu X, Xiong Y: PseUI: pseudouridine sites identification based on RNA sequence information. BMC bioinformatics 2018, 19(1):1-11.
Tahir M, Tayara H, Chong KT: iPseU-CNN: identifying RNA pseudouridine sites using convolutional neural networks. Molecular Therapy-Nucleic Acids 2019, 16:463-470.
Liu K, Chen W, Lin H: XG-PseU: an eXtreme Gradient Boosting based method for identifying pseudouridine sites. Molecular Genetics and Genomics 2020, 295:13-21.
Bi Y, Jin D, Jia C: EnsemPseU: identifying pseudouridine sites with an ensemble approach. Ieee Access 2020, 8:79376-79382.
Lv Z, Zhang J, Ding H, Zou Q: RF-PseU: a random forest predictor for RNA pseudouridine sites. Frontiers in Bioengineering and Biotechnology 2020, 8:134.
Khan SM, He F, Wang D, Chen Y, Xu D: MU-PseUDeep: a deep learning method for prediction of pseudouridine sites. Computational and Structural Biotechnology Journal 2020, 18:1877-1883.
Song B, Tang Y, Wei Z, Liu G, Su J, Meng J, Chen K: PIANO: a web server for pseudouridine-site (Ψ) identification and functional annotation. Frontiers in genetics 2020, 11:88.
Song B, Chen K, Tang Y, Ma J, Meng J, Wei Z: PSI-MOUSE: predicting mouse pseudouridine sites from sequence and genome-derived features. Evolutionary Bioinformatics 2020, 16:1176934320925752.
Li F, Guo X, Jin P, Chen J, Xiang D, Song J, Coin LJ: Porpoise: a new approach for accurate prediction of RNA pseudouridine sites. Briefings in bioinformatics 2021, 22(6):bbab245.
Niu M, Zhang J, Li Y, Wang C, Ma Q: CirRNAPL: A web server for the identification of circRNA based on extreme learning machine. Computational and Structural Biotechnology Journal 2020, 18(869–881).
Lin HL-MZ-HL-XTC: Evaluation of different computational methods on 5-methylcytosine sites identification. Briefings in bioinformatics 2020, 21(3):982-995.
Chai D, Jia C, Zheng J, Zou Q, Li F: Staem5: A novel computational approachfor accurate prediction of m5C site. Molecular therapy Nucleic acids 2021, 26:1027-1034.
Bonidia RP, Machida JS, Negri T, Alves W, Sanches DS: A Novel Decomposing Model With Evolutionary Algorithms for Feature Selection in Long Non-Coding RNAs. IEEE Access 2020(2169-3536):1.
Wang H, Liu H, Huang T, Li G, Zhang L, Sun Y: EMDLP: Ensemble multiscale deep learning model for RNA methylation site prediction. BMC Bioinformatics 2022, 23(1):1-22.
Linder J, Fleur AL, Chen Z, Ljubeti A, Seelig G: Interpreting Neural Networks for Biological Sequences by Learning Stochastic Masks. Nature Machine Intelligence.
Alam W, Tayara H, Chong KT: XG-ac4C: identification of N4-acetylcytidine (ac4C) in mRNA using eXtreme gradient boosting with electron-ion interaction pseudopotentials. Scientific Reports.
Xuan JJ, Sun WJ, Lin PH, Zhou KR, Liu S, Zheng LL, Qu LH, Yang JH: RMBase v2.0: deciphering the map of RNA modifications from epitranscriptome sequencing data. Nucleic Acids Research 2018.
Li J, Huang Y, Yang X, Zhou Y, Zhou Y: RNAm5Cfinder: A Web-server for Predicting RNA 5-methylcytosine (m5C) Sites Based on Random Forest. Scientific Reports 2018, 8(1).
Yhy A, Chi MA, Jsw A, Hui YA, Hui DA, Sgh A, Ywlbc D: Prediction of N7-methylguanosine sites in human RNA based on optimal sequence features. Genomics 2020, 112(6):4342-4347.
Klimo M, Luká P, Tarábek P: Deep Neural Networks Classification via Binary Error-Detecting Output Codes. Applied Sciences 2021, 11(8):3563.
Song ZCZLTM-LLRZRPAIW-CCISJDL: iLearn: an integrated platform and meta-learner for feature engineering, machine-learning analysis and modeling of DNA, RNA and protein sequence data. Briefings in bioinformatics 2020, 21(3):1047-1057.
Zhang L, Qin X, Liu M, Xu Z, Liu G: DNN-m6A: A Cross-Species Method for Identifying RNA N6-methyladenosine Sites Based on Deep Neural Network with Multi-Information Fusion. Genes, 12(3):354.
Tayara H, Tahir M, Chong KT: iSS-CNN: Identifying splicing sites using convolution neural network. Chemometrics and Intelligent Laboratory Systems 2019, 188.
Qiu WR, Jiang SY, Xu ZC, Xiao X, Chou KC: iRNAm5C-PseDNC: identifying RNA 5-methylcytosine sites by incorporating physical-chemical properties into pseudo dinucleotide composition. Oncotarget 2017, 8(25).
Habib U, Rehman HU, Ijaz U, Khan A: Detecting N6-methyladenosine sites from RNA transcriptomes using random forest. Journal of Computational Science 2020, 47(6):101238.
Vacic V, Iakoucheva LM, Radivojac P: Two Sample Logo: a graphical representation of the differences between two sets of sequence alignments. Bioinformatics (Oxford, England) 2006, 22(12):1536.
Lundberg S, Lee SI: A Unified Approach to Interpreting Model Predictions. 2017.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Definer: A computational method for accurate identification of RNA pseudouridine sites based on deep learning

Status:

Version 1

Abstract

Background

Results

Conclusion

Figures

1 Introduction

2 Materials and methods

2.1 The entire experimental framework

2.2 Data collection

2.3 Feature Extraction

2.3.1 One-hot Encoding

2.3.2 NCP

2.4 Classifier

2.4.1 Overall framework of the model(Definer)

2.4.2 CNN

2.4.3 Gate Recurrent Unit

2.4.4 Attention Mechanism

2.5 Performance Evaluations

3 Results and discussion

3.1 Distribution of nucleotide positions at pseudouridine sites

3.2 Performance comparison analysis of different feature extraction methods

3.3 Performance comparison analysis of different models

3.4 Comparative analysis on Independent Datasets

3.5 Comparison with other existing methods

3.6 Software engineering

4 Conclusion

Declarations

Ethics approval and consent to participate

Consent to publication

Availability of data and materials

Competing interests

Funding

Authors’ contributions

Acknowledgments

References

Additional Declarations

Status:

Version 1