A method for predicting subcellular localization of long non-coding RNA based on deep learning
By employing deep learning methods, one-hot encoding, and feature extraction techniques, combined with CNN-Maxpooling and Bi-LSTM models, the problem of low accuracy in lncRNA subcellular localization prediction was solved, achieving higher prediction accuracy and recall.
Patent Information
- Application Number
- CN202310266369.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-13
- Publication Date
- 2025-11-11
- Estimated Expiration
- 2043-03-13
AI Technical Summary
Existing long non-coding RNA (lncRNA) subcellular localization predictors have low accuracy, poor recall, and poor F-measure results in predicting five subcellular localizations: Cytoplasm, Nucleus, Exosome, Ribosome, and Cytoso.
A deep learning-based approach was adopted, using one-hot encoding, electron-ion interaction potential, and nucleotide chemical properties as input features. A network architecture was constructed by combining CNN-Maxpooling, Bi-LSTM, and Attention_LS models to process lncRNA sequences, and the predicted information was output through a fully connected network.
It improves the prediction accuracy of lncRNA subcellular localization, significantly outperforming traditional machine learning methods and existing predictors, especially in the prediction of imbalanced categories.
Smart Images

Figure CN116417066B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bioinformatics, and in particular to a method for predicting the subcellular localization of long non-coding RNAs based on deep learning. Background Technology
[0002] Non-coding RNAs have attracted widespread attention from researchers and are associated with the development of various human diseases. Long non-coding RNAs (lncRNAs) are a class of non-coding RNA molecules (more than 200 nucleotides) transcribed from DNA but not translated into proteins. LncRNAs play important roles in various biological processes, including the regulation of gene expression, alternative splicing, nuclear organization, and genomic imprinting.
[0003] lncRNAs are now considered to play crucial roles in many cellular processes, including the cell cycle, differentiation, and metabolism, as well as disease. Recent evidence also suggests that lncRNAs play a role in viral infection. The past decade has seen a surge in research centered on lncRNAs, which exert their effects through a variety of mechanisms.
[0004] lncRNAs can act as "miRNA sponges" to regulate miRNA levels, thereby influencing the expression of miRNA targets. LncRNAs can regulate transcriptional activity or pathways in response to specific stimuli. Although lncRNA expression levels are generally lower than mRNA levels, they exhibit stronger tissue-specific expression patterns, indicating an indispensable role in cell type-specific processes. Due to the complexity of their molecular functions, lncRNA-related research is receiving increasing attention. Therefore, identifying the subcellular localization of lncRNAs is crucial for understanding their function.
[0005] However, there are few predictors for lncRNA subcellular localization. Existing predictors have low accuracy in predicting the five subcellular localizations of Cytoplasm, Nucleus, Exosome, Ribosome, and Cytoso, and their precision, recall, and F-measure results for each category on the test set are poor. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention provides a method for predicting subcellular localization of long non-coding RNAs based on deep learning. This method solves the technical problems mentioned in the background art, where existing predictors for lncRNA subcellular localization have low prediction accuracy in five subcellular localizations: Cytoplasm, Nucleus, Exosome, Ribosome, and Cytoso. Furthermore, they exhibit poor precision, recall, and F-measure detection results for each category on the test set.
[0007] The above-mentioned technical objective of the present invention is achieved through the following technical solution:
[0008] A method for predicting subcellular localization of long non-coding RNAs based on deep learning includes:
[0009] (1) Construct a benchmark dataset;
[0010] (2) Use one-hot encoding to splice the electron-ion interaction potential (EIIP) and chemical properties (NCP) of each nucleotide as input features;
[0011] (3) Construct a network architecture to process lncRNA;
[0012] (4) Output prediction information through a fully connected network.
[0013] Compared with existing technologies, the present invention has the following advantages in outputting prediction information through a fully connected network:
[0014] 1. The lncRNA sequence was encoded using one-hot encoding, electron-ion interaction potential, and nucleotide chemistry. This characteristic information, after five-fold cross-validation using various modules, showed a slight advantage over the results of other parts.
[0015] 2. CNN-Maxpooling was used to extract information from different combinations of nucleotides. When the kernel size is set to the same as k in k-mer, it is an improved version of k-mer. No paging or phasing is performed during data processing. This method reduces the loss of effective information and the impact of sparse matrices. Multiple convolution operations with different kernel sizes are used to represent different combinations of nucleotides for information extraction.
[0016] 3. The Bi-LSTM model and the Attention_LS model are used together to automatically extract long-range information of lncRNA sequences. The combination of these two data models has good results. Attached Figure Description
[0017] Figure 1A data distribution chart;
[0018] Figure 2 This is a framework diagram;
[0019] Figure 3 AUC plot;
[0020] Figure 4 Confusion matrix diagram for five categories: DeepLncLoc, lncLocator, and LncLST;
[0021] Figure 5 Confusion matrix for iloc_lncRNA, iloc_lncRNA2.0 and LncLST four-class classification. Detailed Implementation
[0022] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0023] Example:
[0024] A method for predicting subcellular localization of long non-coding RNAs based on deep learning includes:
[0025] I. Dataset
[0026] We retrieved known lncRNA subcellular localization information from the RNAlocate database. We collected 42,190 RNA-related subcellular localization entries from RNAlocate. We generated a benchmark dataset using the following procedure to train and test our model.
[0027] (1) Download all 42,190 RNA-related subcellular localization entries from the RNAlocate database.
[0028] (2) 2383 lncRNA subcellular localization entries were selected from 42190 RNA subcellular localization entries.
[0029] (3) Some lncRNAs have multiple entries; we merged these entries with the same gene name. Then, we removed lncRNAs that did not have sequence information in NCBI and Ensembl.
[0030] (4) Because most lncRNAs have only one subcellular location, we selected lncRNAs that are located at only one location for model building in our study.
[0031] (5) The filtered dataset covers seven different subcellular localizations. Two of the seven subcellular localizations have very few samples (less than 10). Therefore, we removed lncRNAs with insufficient subcellular localization samples.
[0032] (6) We use cd-hit to remove sequence information with a sequence similarity of over 80%.
[0033] Finally, we constructed a benchmark dataset of 842 lncRNAs covering five subcellular localizations, including Nucleus, Exosome, Cytoplasm, Cytosol, and Ribosome.
[0034] The processed dataset distribution is as follows Figure 1 As shown in the figure, the lengths of lncRNAs distributed in Cytoplasm, Cytosol, and Exosome are mostly concentrated between 200 nt and 3000 nt, while the lengths of lncRNAs located in Nucleus and Ribosome are concentrated between 200 nt and 3000 nt and greater than 12000 nt.
[0035] II. Feature Coding.
[0036] We use one-hot encoding to concatenate the electron-ion interaction potential (EIIP) and nucleotide chemical properties (NCP) of each nucleotide as part of the input features. Each nucleotide is encoded by a four-bit binary vector, with nucleotides A, T, C, and G set to [1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], and [0, 0, 0, 1].
[0037] The four EIIP values were assigned as A: 0.1260, C: 0.1340, G: 0.0806, and T: 0.1335. The EIIP encoding of RNA directly uses the EIIP value representing the nucleotides in the DNA sequence.
[0038] Nucleotide chemical properties (NCP): There are four different types of nucleotides in RNA sequences, each with different chemical structures and binding properties. Based on ring structures, functional groups, and hydrogen bonds, nucleotides in lncRNA sequences can be divided into three categories. C and T have monocyclic structures, while A and G have bicyclic structures. A and C belong to amino groups, while G and T belong to ketone groups. A and T form strong hydrogen bonds, while C and G form weak hydrogen bonds. Combining these chemical characteristics, the following equation is used to represent the i-th nucleotide in the RNA sequence:
[0039]
[0040]
[0041]
[0042] Based on their chemical properties, A can be coded as (1, 1, 1), C as (0, 1, 0), G as (1, 0, 0), and T as (0, 0, 1).
[0043] Combining the three features, the following equation is used to represent the form of each lncRNA.
[0044]
[0045] For performance evaluation, a 5-fold cross-validation dataset was constructed using a non-redundant dataset. Specifically, the non-redundant dataset was divided into 5 folds, each with a similar distribution of subcellular localization categories. The 5-fold cross-validation dataset can be directly used for comparison with other methods. All data (including benchmark datasets for all lncRNAs and independent test datasets) are freely available on the LncLST web server.
[0046] III. Network Architecture
[0047] The framework of LncLST is as follows Figure 2 As shown, our model can accept lncRNA sequences of different lengths. Input sequences of different lengths can be encoded into features of shape L*8. The feature shape of lncRNA sequences input into the neural network of different lengths is batch*L*8, where batch is the number of data samples captured in one training session, and L is the length of the lncRNA sequence input each time. To avoid information loss, we do not truncate or pad the lncRNA length; the length of the lncRNA input to the neural network is variable. We use the CNN1d-Maxpooling module to extract information from different combinations of nucleotides. After this module, the feature dimension of the input sequence changes from batch*L*8 to batch*128*8, and outputs two types of features with convolution kernels of 3 and 5. Next, we input both types of features into the BI-LSTM module and the Attention_LS module, respectively. Both models can effectively extract the information contained in long sequences. After processing by the long sequence feature extraction module, four-way features are output. Then, we concatenate the features according to different kernels, changing the feature dimension to batch*256*8, and then input them into the Text-CNN module. In Text-CNN, the feature with the largest value in a sequence is extracted. Finally, we concatenate the feature information from the two Text-CNN inputs and input them into a fully connected network. The fully connected network contains 5 neurons, corresponding to 5 localization categories. After passing through the SoftMax function, the probability of each category is between 0 and 1, and the sum of the probabilities of multiple categories is 1.
[0048] IV. CNN1d-Maxpooling
[0049] In most cases, the length of a sequence is often achieved by zero-padding and truncation to make each sequence the same length. However, this can result in excessively long sequences or the loss of some information, which can affect the feature extraction of neural networks and reduce accuracy.
[0050] After encoding, the feature dimension of the input CNN1D-Maxpooling is batch*L*8. To achieve efficient extraction of LncRNA sequence information using one-dimensional convolution operations and better results than k-mer, we swapped the last two dimensions of the input sequence, changing the shape to batch*8*L. During one-dimensional convolution, the convolution kernel slides at the last dimension, thereby extracting information from different combinations of nucleotides. To more effectively extract information from different combinations of nucleotides, we performed three CNN+Maxpooling operations, with kernels set to 3 and 5 for each operation. In the first operation, we set the number of CNN filters to 64. After the first convolution operation, the feature dimension becomes batch*64*(L-kernel+1). Then, we performed adaptive max pooling, changing the feature dimension to batch*64*512. In the second operation, we set the number of filters to 64 and set the last dimension of the adaptive pooling to 256, thus changing the feature dimension to batch*64*256. In the third operation, we set the number of filters to 32 and the adaptive pooling dimension to 128, resulting in both sizes of kernels having a dimension of batch*32*128. We then swapped the second and third dimensions for subsequent training, changing the feature dimension to batch*128*32.
[0051] V. Sequence Information Extraction Module
[0052] Transformers have achieved success in both language and vision domains. However, scaling them to long sequences (e.g., long documents or high-resolution images) is prohibitively expensive due to the quadratic time and memory complexity of self-attention mechanisms relative to the length of the input sequence. Long-Short Transformer (Transformer-LS) is an efficient self-attention mechanism for modeling long sequences with linear complexity for language and vision tasks. It combines a novel long-term attention and dynamic projection to model long-range relevance, with a short-term attention to capture fine-grained local relevance. Transformer-LS uses a dual normalization strategy to address the scale mismatch between the two attention mechanisms. Transformer-LS can be applied to autoregressive and bidirectional models without increasing complexity. In this paper, we primarily use the attention mechanism of Transformer-LS because selecting the entire Transformer-LS model is prone to overfitting, leading to poor training results.
[0053] LSTM stands for Long Short-Term Memory, a type of RNN (Recurrent Neural Network). Due to its design characteristics, LSTM is well-suited for modeling time-series data, such as text data. BiLSTM, short for Bi-directional Long Short-Term Memory, is a combination of forward and backward LSTM. However, using LSTM to model sequences presents a problem: it cannot encode information from back to front. BiLSTM can better capture bidirectional sequence information.
[0054] We input the features from the CNN-Maxpooling module into LongShortAttention and BiLSTM. Considering that too many attention heads could lead to overfitting, we set the number of attention heads in LongShortAttention to 6 and the dropout to 0.65. By controlling the window size w of Short-Term Attention to 128 and the projection dimension r of Long-Term Attention to 4, we project the 128 feature length onto a lower dimension of 4, significantly reducing space complexity. We set the hidden layer size of BiLSTM to 16 and the dropout to 0.5. Then, we concatenate the matrix output by LongShortAttention with the matrix output by BiLSTM, that is, concatenating the LongShortAttention feature B*128*32 and the BiLSTM feature B*128*(16*2) into B*128*64, thus achieving feature integration.
[0055] VI. Text CNN Module
[0056] Text CNN is a powerful deep learning network architecture for text classification. Traditional CNNs are 2D CNNs designed for processing 2D image data. In reality, text can be viewed as a 1D image, so we can use a 1D CNN to extract text features. We use Text CNN and fully connected operations to predict the subcellular localization of lncRNAs. Text CNN uses 1D convolutional layers and max-pooling layers to extract sequence features. Specifically, the entire sequence can be represented by an L*D matrix. The lncRNA sequence can be viewed as a 1D image with a width of D, a height of 1, and L channels. Text CNN uses three convolutional kernels (sizes = 1, 3, 5) to further aggregate long-range nucleotide features, and then uses max-pooling to obtain the most salient features across the entire sequence. The vectors output from the max-pooling layers are concatenated and used as input to a fully connected layer with a SoftMax function to perform the final prediction.
[0057] VII. Evaluation Indicators
[0058] Similar to previous studies, we used accuracy (ACC), macro F-measure, and area under the ROC curve (AUC) as evaluation metrics to assess LncLST and other methods in the study.
[0059]
[0060]
[0061]
[0062]
[0063] Among them TP (i) FP (i) FN (i) denoted as true positive, false positive, and false negative for class i, respectively, and m represents the number of subcellular localizations of LncRNA.
[0064] VIII. Implementation Details
[0065] LncLST is implemented using PyTorch, and we use focal loss as the loss function. Focal loss was originally proposed by Kaiming He and was initially used in the image domain to address model performance issues caused by imbalanced data. This paper uses focal loss to handle the imbalanced data problem.
[0066]
[0067] To avoid overfitting, we set the dropout for convolutional and fully connected operations to 0.5, and the attention dropout to 0.65. Finally, we trained LncLST using AdamW with an initial learning rate of 0.0002 and weight decay of 0.001.
[0068] Results and Discussion
[0069] 1. Hyperparameter optimization for LncLST
[0070] We used the results of 5-fold cross-validation to fine-tune the hyperparameters of LncLST. In our model, numerous hyperparameters influence the prediction results, such as the kernel size, the number of kernels, the pooling layer size, the number of attention heads, and the initial learning rate. In our study, we primarily focused on the kernel size K, the number of kernels S, the pooling layer size P, and the number of attention heads N. We used a grid search algorithm to find the optimal hyperparameters. We selected the kernel size in [3, 4, 5, 6, 7], the number of kernels in [32, 64, 128], the pooling layer size in [64, 128, 256], and the number of attention heads in [2, 4, 6, 8]. Ultimately, we found that the training performance was optimal when the kernel size was 3+5, the number of kernels was 32, the pooling layer size was 128, and the number of attention heads was 6. In this case, with five-fold cross-validation, LncLST has ACC, MAF, and AUC values of 0.594, 0.455, and 0.841, respectively.
[0071] 2. Comparison with traditional machine learning methods under different k-mer conditions
[0072] Because k-mer kernels are widely used as feature inputs in traditional machine learning methods, we compare LncLST with four traditional machine learning methods: SVM, RF, LR, and simple NN. We implemented all machine learning models in the Python scikit-learn (v 0.21.1) library. For SVM, we used the rbf kernel. For LR, RF, and NN, we used the default parameters from scikit-learn. Considering that our LncLST uses a combination of 3+5 kernels, we chose k as {3, 5, 3+5} in our comparison with traditional machine learning methods. The results are shown in Table 1.
[0073] As can be seen from Table 1, when using the combination of k=3+k=5, all machine learning methods achieve good results. In addition, LncLST is significantly better than other machine learning methods in all evaluation metrics.
[0074] Table 1
[0075]
[0076] from Figure 3ROC curves for other machine learning methods were plotted. The graph shows that LncLST outperforms other machine learning methods in all five subcellular localizations. The leading results are most significant in cytoplasm and ribsome, while it slightly outperforms NN methods in exosome. In summary, these results indicate that LncLST is superior to traditional machine learning methods.
[0077] (1) Ablation test
[0078] To demonstrate the crucial role each module plays in our model's predictions, we conducted ablation experiments on LncLST using various modules and performing five-fold cross-validation. The results are shown in Table 2. The results indicate that the fully-structured LncLST performs best in five-fold cross-validation. However, using one-hot encoding as input features only slightly outperforms the results of ablation. In the experiment using k-mer+word2vec, the results are significantly lower than the other experiments, highlighting the important role of the CNN1d-MaxPooling module. Replacing the conventional k-mer+word2vec structure with CNN1d-MaxPooling yields a significant advantage.
[0079] Table 2
[0080]
[0081] One-hot encoding means that only one-hot encoding is used as the input feature, without using electron-ion interaction potential (EIIP) and nucleotide chemical properties (NCP). That is, the feature dimension of each nucleotide is reduced from 8 dimensions to 4 dimensions.
[0082] k-mer+word2vec means replacing the input features of LncLST and the CNN1d-MaxPooling module with the regular k-mer+word2vec structure, and truncating or padding the input sequence to determine the length of the input sequence as 8000nt.
[0083] Only Attention_LS means removing the BiLSTM structure and using only Only Attention_LS.
[0084] Only BiLSTM means removing the Attention_LS structure and using only BiLSTM.
[0085] (2) Comparison with existing predictors
[0086] In our previous experiments, we applied five-fold cross-validation. To further evaluate the performance of LncLST in predicting subcellular localization of lncRNAs, we compared LncLST with existing predictors using an independent test set. The current predictors we selected met the following criteria: (i) availability of a web server or standalone version; (ii) requiring only the lncRNA sequence as input; and (iii) output including subcellular localization prediction scores. Therefore, lncLocator, iLoc-lncRNA, iloc-lncRNA2.0, and DeepLncLoc met these criteria. LncLocator can predict five subcellular localizations of lncRNAs, including Cytoplasm, Nucleus, Exosome, Ribosome, and Cytosol. iLoc-lncRNA and iLoc-lncRNA2.0 can predict four subcellular localizations of lncRNAs, including Cytoplasm, Nucleus, Exosome, and Ribosome. We used lncLocator (predictors available at http: / / www.csbio.sjtu.edu.cn / bioinf / lncLocator / ), iLoc-lncRNA (predictors available at http: / / lin-group.cn / server / iLoc-LncRNA / predictor.php), iLoc-LncRNA2.0 (predictors available at http: / / lin-group.cn / server / iLoc-LncRNA(2.0) / predictor.php), and DeepLncLoc (predictors available at http: / / bioinformatics.csu.edu.cn / DeepLncLoc / ). We compared LncLST with the four predictors (lncLocator, iLoc-lncRNA, iLoc-lncRNA2.0, and DeepLncLoc) using an independent test set. The test set was a combination of samples selected from another lncRNA subcellular localization database, lncSLdb, and the Allexperimental RNA subcellular localization data dataset of RNAlocate2.0. Because the lncSLdb database only collects five subcellular localizations—nucleus, chromosome, cytoplasm, nucleoplasm, and ribosome—and does not record cytosol and exosome, we randomly selected some samples from lncSLdb.The RNAlocate2.0 dataset contains a large number of multi-label lncRNAs; we filtered out lncRNAs with only single labels. We integrated the two datasets and then used the cd-hit tool to remove redundant sequences with a cutoff of 80%. The final independent test set contained 91 nucleus samples, 80 cytoplasm samples, 46 exosome samples, 17 cytosol samples, and 10 ribosome samples. Since iLoc-lncRNA and iloc-lncRNA2.0 treat cytoplasm and cytosol as a single class, they only predict four classes: cytoplasm, nucleus, exosome, and ribosome. For fairness, when comparing LncLST with iloc-lncRNA and iloc-lncRNA2.0, we also treat cytoplasm and cytosol as a single class. The confusion matrices of DeepLncLoc, lnclocator, and LncLST for the independent test set are shown below. Figure 4 As shown.
[0087] The confusion matrices of iloc_lncRNA, iloc_lncRNA2.0, LncLST, and LncLST for independent test sets are as follows: Figure 5 As shown. Figure 5 In the table, each row represents the true class, each column represents the predicted class, and the diagonal elements represent the number of correctly predicted samples. In 5-class classification, our LncLST correctly predicted 113 out of 244 LncRNAs, slightly higher than DeepLncLoc's 92 and significantly higher than lncloctor's 82. In 4-class classification, our LncLST correctly predicted 122 out of 244 LncRNAs, slightly higher than iloc-LncRNA's 118, but significantly lower than iloc_LncRNA2.0's 102. The results for DeepLncLoc, lncLocator, iLoc-lncRNA, iLoc-LncRNA2.0, and LncLST are shown in Table 3.
[0088] Clearly, LncLST is more accurate than lncLocator and DeepLncLoc. DeepLncLoc (class 5) has macroprecision, macro recall, and macro F-measure values of 0.491, 0.384, and 0.397, respectively, significantly higher than lncLocator (0.353, 0.325, and 0.297) and slightly higher than DeepLncLoc (0.412, 0.373, and 0.345). Similar results were observed when we compared LncLST (class 4) with iLoc-lncRNA and iloc-lncRNA2.0.
[0089] All results indicate that LncLST can serve as a useful tool for predicting the subcellular localization of lncRNAs.
[0090] Table 3
[0091] Comparison of prediction performance of LncLST, lncLocator, iLoc-lncRNA, iloc-lncRNA2.0, and DeepLncLoc on the test set
[0092]
[0093]
[0094] Table 4 reports the precision, recall, and F-measure of LncLST, lnclocator, and deepplncloc for each class on the test set. We observed that LncLST's F-measure for Cytoplasm was slightly lower than lnclocator's, but significantly higher than deepplncloc's. We also noted that lnclocator performed very poorly for Ribosomes. One possible explanation is that lnclocator has too many cytoplasmic samples, causing the machine learning model to favor Cytoplasm and resulting in poor performance for other classes. For classification of small samples such as Ribosomes and Cytosols, LncLST achieved the highest F-measure score, which to some extent indicates that LncLST has a certain improvement over lnclocator and deepplncloc models in handling imbalanced data.
[0095] Table 4 shows the precision, recall, and F-measure of LncLST, lnclocator, and deepplncloc for each class on the test set.
[0096]
[0097] Table 5 reports the precision, recall, and F-measure of LncLST, iLoc-lncRNA, and iLoc-lncRNA2.0 for each class on the test set. As can be seen from the table, LncLST's performance in each category is significantly higher than that of iLoc-lncRNA2.0. When predicting Ribosomes, LncLST's accuracy is significantly higher than that of iLoc-lncRNA and iLoc-lncRNA2.0 models. Furthermore, the F-measure of LncLST for each category is not significantly different, indicating that LncLST has a significant advantage in predicting imbalanced classes.
[0098] Table 5
[0099] LncLST and iLoc-lncRNA, iloc-lncRNA2.0 precision, recall, and [other parameters] for each class on the test set.
[0100] F-measure
[0101]
[0102] 9. LncLST's webpage
[0103] We provide a free web server (lncLST) for predicting the subcellular localization of lncRNAs. LncLST requires an lncRNA sequence greater than 200 nt but less than 50,000 nt. Users then click the predict button to view the prediction results. The prediction results for each category are displayed in a table, and the predicted category is output and shown on the screen. Typically, LncLST takes less than 2 seconds to predict the subcellular localization of lncRNAs.
[0104] X. Conclusion
[0105] In this study, we propose an open-source deep learning model for predicting the subcellular localization of lncRNAs. Unlike previous computational methods that use k-mer features to encode lncRNA sequences, LncLST offers an improved approach to extracting surrounding nucleotide information and utilizes attention_LS and BiLSTM to handle long sequence features. LncLST features three novel design elements compared to previous research. Our extensive results demonstrate that LncLST outperforms all traditional machine learning models with different k-mer features and state-of-the-art predictors. We believe LncLST can serve as a useful tool for predicting the subcellular localization of lncRNAs.
[0106] 1. We used one-hot encoding, electron-ion interaction potential, and nucleotide chemistry to encode lncRNA sequences. This characteristic information showed good results in ablation experiments.
[0107] 2. We used CNN-Maxpooling to extract information from different combinations of nucleotides. When the kernel size is set to the same as k in k-mer, it is an improved version of k-mer. Our method does not perform any stages or padding during data processing, which reduces the loss of effective information and the impact of sparse matrices. We can use multiple convolution operations with different kernel sizes to represent different combinations of nucleotides for information extraction.
[0108] 3. We use a combination of Bi-LSTM and Attention_LS models to automatically extract long-range information from lncRNA sequences. The combination of these two data models has good results.
[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for predicting subcellular localization of long non-coding RNAs based on deep learning, characterized in that, include: (1) Construct a benchmark dataset; (2) Use one-hot encoding to splice the electron-ion interaction potential (EIIP) and chemical properties (NCP) of each nucleotide as input features; (3) Construct a network architecture to process lncRNA; the specific process of step (3) is as follows: I. Replace the commonly used k-mer with one-dimensional convolution. The specific process of step one is as follows: The feature dimension of the input CNN1D-Maxpooling is batch*L*8. The last two dimensions of the input sequence are swapped, resulting in a shape of batch*8*L. Three CNN+Maxpooling operations are performed, with the kernels set to 3 and 5 for each operation. In the first operation, the number of CNN filters is set to 64. After the first convolution operation, the feature dimension becomes batch*64*(L-kernel+1), and then... The adaptive max pooling operation transforms the feature dimension into batch*64*512. In the second operation, the number of filters is set to 64, and the dimension of the last dimension after adaptive pooling is set to 256, thus transforming the feature dimension into batch*64*256. In the third operation, the number of filters is set to 32, and the adaptive pooling dimension is set to 128, resulting in both sizes of kernels having a dimension of batch*32*128. The second and third dimensions are then swapped for subsequent training, transforming the feature dimension into batch*128*32.
2. Use the long-short Transformer framework and the Bi-LSTM framework to process sequence information; the specific process of step 2 is as follows: use the Attention mechanism of Transformer-LS; input the features passed through the CNN-Maxpooling module into LongShortAttention and BiLSTM; Set the number of attention heads in LongShortAttention to 6; set the dropout of attention to 0.65; set the window w of Short-Term Attention to 128; set the projection dimension r of Long-Term Attention to 4, that is, project the feature length of 128 to a lower dimension of 4; set the hidden layer size of BiLSTM to 16; set the dropout to 0.5; concatenate the matrix output by LongShortAttention with the matrix output by BiLSTM, that is, concatenate the LongShortAttention feature B*128*32 and the BiLSTM feature B*128*(16*2) into B*128*64, thereby realizing feature integration. Third, use TXT CNN to reduce the dimensionality of sequence information; the specific process of step three is as follows: use an L*D matrix to represent the entire sequence; the lncRNA sequence is represented as a one-dimensional image with a width of D, a height of 1, and a channel of L; Text CNN uses three convolutional kernels with sizes of 1, 3, and 5 to aggregate long-distance nucleotide features again, and then uses max pooling to obtain the most significant features on the entire sequence; The vectors output by the max pooling layer are concatenated and used as input to a fully connected layer with a SoftMax function to perform the final prediction. (4) Output prediction information through a fully connected network.
2. The method for predicting subcellular localization of long non-coding RNA based on deep learning according to claim 1, characterized in that, Step (1) specifically involves downloading all 42,190 RNA-related subcellular localization entries from the RNAlocate database; Select lncRNA subcellular localization entries; The selected lncRNAs were processed; A benchmark dataset of 842 lncRNAs was constructed, covering five subcellular localizations, including Nucleus, Exosome, Cytoplasm, Cytosol, and Ribosome.
3. The method for predicting subcellular localization of long non-coding RNA based on deep learning according to claim 1, characterized in that, Step (2) is as follows: set the nucleotides A, T, C, and G to [1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]; The EIIP values of nucleotides in a DNA sequence are represented by the EIIP codes of RNA, where the four EIIP values are A: 0.1260, C: 0.1340, G: 0.0806 and T: 0.1335. The nucleotides in a lncRNA sequence can be divided into three categories, and the i-th nucleotide in the RNA sequence can be represented by the following equation: Based on chemical properties, A is encoded as (1, 1, 1), C as (0, 1, 0), G as (1, 0, 0), and T as (0, 0, 1). Combining the three features, the following equation is used to represent the form of each lncRNA.
4. The method for predicting subcellular localization of long non-coding RNA based on deep learning according to claim 1, characterized in that, The specific process of step (4) is as follows: the user inputs the LncRNA sequence on the web server; Click the prediction button; The prediction results for each category are displayed in a table, and the predicted category is output and shown on the screen.
Citation Information
Patent Citations
Deep learning based long-chain non-coding RNA subcellular position prediction algorithm
CN107577924A
Multi-scale CNN-BiLSTM non-coding RNA interaction relationship prediction method with introducing attention
CN111341386A