Skip to main content
Download PDF

CPPVec: an accurate coding potential predictor based on a distributed representation of protein sequence

BMC Genomics volume 24, Article number: 264 (2023) Cite this article

Abstract

Long non-coding RNAs (lncRNAs) play a crucial role in numbers of biological processes and have received wide attention during the past years. Since the rapid development of high-throughput transcriptome sequencing technologies (RNA-seq) lead to a large amount of RNA data, it is urgent to develop a fast and accurate coding potential predictor. Many computational methods have been proposed to address this issue, they usually exploit information on open reading frame (ORF), protein sequence, k-mer, evolutionary signatures, or homology. Despite the effectiveness of these approaches, there is still much room to improve. Indeed, none of these methods exploit the contextual information of RNA sequence, for example, k-mer features that counts the occurrence frequencies of continuous nucleotides (k-mer) in the whole RNA sequence cannot reflect local contextual information of each k-mer. In view of this shortcoming, here, we present a novel alignment-free method, CPPVec, which exploits the contextual information of RNA sequence for coding potential prediction for the first time, it can be easily implemented by distributed representation (e.g., doc2vec) of protein sequence translated from the longest ORF. The experimental findings demonstrate that CPPVec is an accurate coding potential predictor and significantly outperforms existing state-of-the-art methods.

Peer Review reports

Introduction

Recently, long non-coding RNAs (lncRNAs, \({> {200}}\)nt) have received increasingly more attention for their participation in numerous important biological processes (e.g., gene regulation and expression [1], cell cycle regulation [2]). The mutations and dysregulations in lncRNAs can cause human diseases, such as cancer, cardiovascular and neurodegenerative diseases [3,4,5,6]. It is still a challenging task to distinguish lncRNAs from messenger RNAs (mRNAs), this is because 1) they often have very similar features, such as poly(A) tails, splicing and approximate sequence length [7]; 2) lncRNAs may contain small open reading frame (sORF) that encodes micropeptides [8], which could induce false positives; 3) there are considerable indel errors [9] during the process of sequencing and assembly.

Many computational methods have been proposed to distinguish lncRNAs from mRNAs in the past years [10,11,12,13]. These methods mainly exploit five kinds of information: 1) open reading frame (ORF). The longest ORF of an RNA sequence is often extracted because it is likely to be the correct ORF where a protein is translated [14], then the ORF length, ORF integrity and ORF coverage are selected as ORF features that are effective and widely used by current methods. CPAT [12] identified that ORF length is the most important feature for coding potential prediction. However, ORF features are more likely to be correct when no sequencing or assembly errors occur, and hence are not suitable for platforms with indel errors, e.g., Roche (454) [15]. 2) protein sequence. The physicochemical properties of the protein sequence translated from the longest ORF can also carry information for coding potential prediction. CPC2 [11] used isoelectric point, and CPPred adds the other two properties (e.g., gravy and instability) mentioned by CPC2. 3) k-mer (e.g., codon usage (3-mer), hexamer usage (6-mer)). k-mer features are often calculated by counting the frequency of fixed-length words (k-mer) that occur in an RNA sequence, or using its variant, e.g., usage frequency of adjoining nucleotide triplets (ANT) in CNCI [16]. k-mer features are effective, and even robust (overlapping k-mer in PLEK [10]) for coding potential prediction for the fact that the distribution over k-mer is significantly different in mRNAs to lncRNAs. Recent study [17] also combined k-mer features directly with deep neural networks (e.g., Convolutional Neural Network (CNN)) to identify lncRNAs from mRNAs and achieved better performance than traditional classifiers (e.g., Support Vector Machine (SVM)). Despite the effectiveness of k-mer features, they count the occurrence frequencies of continuous nucleotides (k-mer) in the whole RNA sequence, which cannot reflect local contextual information of k-mer. Moreover, the increase of k leads to a very long and sparse vector representation, which not only induce noise, but also computational burden in real cases [18]. 4) Evolutionary signatures. This information is based on the sequence conservation that RNAs belonging to the same class often have similar sequence composition (e.g., base composition, transition, motifs) during the evolutionary process. CONC [19] uses amino acid composition and sequence entropy. CPPred employs CTD (composition (C), transition (T) and distribution (D)) features [20], they indicate that CTD features are particularly important for coding potential prediction of sORF. However, evolutionary signatures (e.g., CTD features) that these methods use are also simple statistics calculated with the continuous nucleotides, which loses contextual information of RNA sequences. 5) Homology information. This information is exploited by alignment-based methods (e.g., CPC [21], PhyloCSF [22]), which performs sequence alignments to known protein database (e.g., UniProt [23]) or well-annotated reference genome to assess the coding potential of transcript. However, these methods heavily depend on sequence alignments, which is not only computationally expensive, but also not suitable for species without known protein database or well-annotated reference genome [10, 16].

Based on the above analysis, here, we explored how to exploit the contextual information of RNA sequence to enhance the performance of coding potential prediction. We developed an accurate coding potential predictor, CPPVec, which exploits the contextual information of RNA sequence based on distributed representation (e.g., doc2vec [24]) of protein sequence translated from the longest ORF. Tests on human, mouse, zebrafish, fruit fly and Saccharomyces cerevisiae datasets demonstrate that CPPVec significantly outperforms existing state-of-the-art methods. To our best knowledge, this is the first attempt to introduce distributed representation to coding potential prediction. There are two main contributions of our proposed method:

  • We exploited the contextual information of RNA sequence for coding potential prediction for the first time, which was easily implemented by using a distributed representation (e.g., doc2vec) of protein sequence. The experimental results demonstrated the effectiveness of distributed representation for coding potential prediction.

  • We fixed hexamer score by calculating it with the first reading frame of the longest ORF instead of the RNA sequence in CPPred and verified the effectiveness of this fixed feature.

The source code and the dataset used in the paper are publicly available at: https://github.com/hgcwei/CPPVec.

Materials and method

Datasets

In this study, we adopted the datasets strictly selected by CPPred to test our proposed method. Two models are built for coding potential prediction, including Human-Model and Integrated-Model. For Human-Model, 50,040 human (Homo sapiens) mRNAs are downloaded from NCBI RefSeq [25] (https://ftp.ncbi.nih.gov/) and 37,297 human ncRNAs are downloaded from Ensembl database [26], released in 26 November 2017 (https://ftp.ensembl.org/). 33360 mRNAs and 24163 ncRNAs are randomly selected as training set (Human-Training), 8557 mRNAs and 8241 ncRNAs are selected as testing set (Human-Testing) after redundancy removal by using CD-hit [27] with sequence identity cutoff \(\ge 80\)%. Moreover, mouse (Mus musculus), zebrafish (Danio rerio), fruit fly (Drosophila melanogaster), S. cerevisiae are also selected as testing sets (e.g., Mouse-Testing, Zebrafish-Testing, S. cerevisiae-Testing, Fruit-fly-Testing) to compare the cross-species prediction performance of different classification methods. They are constructed following the same building strategy as Human-Testing. For Integrated-Model, in order to evade the problem caused by the specificity of species, several species (e.g., human, mouse, zebrafish, fruit fly, S. cerevisiae, nematode (Caenorhabditis elegans) and thale cress (Arabidopsis thaliana)) are downloaded from NCBI RefSeq, including 525,316 mRNAs and 55,198 ncRNAs. To evade the problems of computational burden and data imbalance, 52,530 mRNAs and 27,600 ncRNAs are randomly selected as training set (Integrated-Training), 13,903 mRNAs and 13,903 ncRNAs are randomly selected as testing set (Integrated-Testing) after redundancy removal using CD-hit with sequence identity cutoff \(\ge 80\)%.

Moreover, in order to verify the effectiveness of CPPVec to find novel lncRNAs, we constructed a testing set from EVlncRNAs [28], which is a comprehensive, manually curated and high-quality lncRNAs database validated by low-throughput experiments (e.g., qRT-PCR, knockdown, etc.). We downloaded all the available sequences of lncRNAs and got 37 novel lncRNAs (https://www.sdklab-biophysics-dzu.net/EVLncRNAs2/), then CD-hit with sequence identify cutoff \(\ge 80\)% is used to remove the lncRNAs that are similar to Integrated-Training. Finally, 34 lncRNAs are remaining and selected as an independent testing set.

Distributed representation of protein sequence

Representation learning plays an important role in machine learning methods [29]. A proper representation usually achieves good result for a machine learning task. In the past years, distributed representation has been proved to be a successful data representation approach in natural language processing. Compared with one-hot encoding, distributed representation contains more semantic information about language context and more suitable for tasks such as sentiment classification [30], text classification [31]. Indeed, biological sequences (e.g., DNA, RNA and protein sequences) have many similar characteristics with natural language. For example, they are both symbol sequences that elements in the sequence are arranged in a specified order, on the other hand, they contain a lot of semantic information, many biologists believe that biological sequences are not merely one-dimensional string of symbols, but encode a lot of useful information about molecular structure and functions in themselves [32]. Hence, it is a natural idea to introduce distributed representation in natural language processing to biological sequence analysis. It is firstly introduced by ProtVec [33] to protein family classification and a prediction accuracy of 99% is achieved, then it is pervasive in a wide range of applications for biological sequences analysis, e.g., protein secondary structure prediction [34], RNA-protein binding sites prediction [35, 36].

In this paper, we introduce the distributed representation to coding potential prediction for RNA sequence. To attain this goal, we are faced with three problems: 1) Which kind of sequence should we choose to encode, RNA sequence, the longest ORF extracted from RNA sequence, or protein sequence translated from the longest ORF? 2) How to build a corpus from the chosen sequences? and 3) How to train the corpus and get a distributed representation for each sequence? In our opinion, our application is concerned with coding potential of RNA sequence, and hence we should pay more attention to protein sequence. Moreover, just as a word in natural language, the basic unit of a protein is “word” called codon (corresponding to acid amine), and hence we consider the distributed representation of protein sequence translated from the longest ORF which is more likely to be the correct ORF than other ORFs, we employ the popular framework, doc2vec to generate a vector representation (embedding) of a protein sequence. To be specific, for all the translated protein sequences, we first adopt the following splitting strategy to generate a “document” for each protein sequence:

$$\begin{aligned}{} & {} MNFLLSWVHWSLALLLYL\ldots \rightarrow \\{} & {} MNF, LLS, WVH, WSL, ALL, LYL,\ldots \end{aligned}$$

where a protein sequence is split in a non-overlapping manner with word length of 3. Second, every split protein sequence is formed into a “document” and appended to a corpus, then we can use distributed memory model of paragraph vectors (PV-DM) (Fig. 1) to train the corpus and generate distributed representation for each “document”. In PV-DM, a protein sequence ID and the context of the central “word” (e.g., WVH) are mapped into a unique vector in themselves, which are concatenated together to predict the central “word”. It is inspired by the idea that apart from contextual information, the paragraph vector is also asked to contribute to the prediction task of the central “word”.

Fig. 1
figure 1

The distributed memory model of paragraph vectors (PV-DM) for protein sequence. The trained vector representations of protein sequence and each word contain contextual information of protein sequence

Full size image

In what follows, \(\varvec{s}=s_{1}s_{2}...s_{n}\) denote a “document” generated from a protein sequence. \(s_{i}\) denote the i-th “word” in the “document”, \(\varvec{W}\) is the linear mapping matrix for each “word”, \(\varvec{v}\) is the paragraph vector to be trained, the predicted occurrence probability of the central “word” \(s_{t}\) given its context can be represented as:

$$\begin{aligned} \hat{y} = b + \varvec{U}f(s_{t-k},...,s_{t-1},s_{t+1},...,s_{t+k};\varvec{W},\varvec{v}) \end{aligned}$$
(1)

where \(\varvec{U},b\) are the softmax parameters [24] and 2k is the length of context. f is function that concatenate mapped word vector \(\varvec{W}\cdot s_{i}\) with paragraph vector \(\varvec{v}\). By the concatenated part, the words that have similar context will have similar distributed representations. Note that all the words in a split protein sequence share the same protein sequence ID and paragraph vector. After training with stochastic gradient descent, the generated vector representation (embedding) of the split protein sequence carry contextual information of RNA sequence and can be used for coding potential prediction.

It is worth noting that in our recent paper [37], we use one-hot encoding to capture contextual information of biological sequence for protein coding regions prediction, however, it is not suitable for coding potential prediction for two reasons: 1) protein sequence translated from the longest ORF has a variable length but most of machine learning methods only receive a fixed-length input. 2) one-hot encoding is too low-level to reflect high-level semantic information of biological sequence. Distributed representation elegantly alleviates the above problems, e.g., doc2vec not only naturally converts a variable-length sequence to a fixed-length vector, but also contains a lot of contextual information of RNA sequence.

Performance evaluation of CPPVec

To evaluate the performance of CPPVec, we use the standard performance metrics, such as sensitivity (SN), specificity (SP), accuracy (ACC), precision (PRE), F-score, AUC and MCC. These metrics can be calculated as follows:

$$\begin{aligned}{} & {} SN = \frac{TP}{TP+FN}\\{} & {} SP = \frac{TN}{FP + TN}\\{} & {} PRE = \frac{TP}{TP + FP}\\{} & {} ACC = \frac{TP+TN}{TP+TN+FP + FN}\\{} & {} F-score = \frac{2 * PRE * SN}{PRE + SN}\\{} & {} MCC = \\{} & {} \frac{TP * TN - FP * FN}{(TP+FN)*(TP+FP) * (TN+FP) * (TN +FN)} \end{aligned}$$

All the above metrics are based on the notions of TP, FP, TN, and FN, which correspond to number of positive samples identified correctly, negative samples identified incorrectly, negative samples identified correctly, and positive samples identified incorrectly, respectively. The MCC is an overall measurement of performance and another objective assessment index. AUC is the area under the receiver operating characteristic curve, it can be calculated by using the trapezoidal areas created between each ROC points.

Results and discussion

Pipeline of CPPVec

The pipeline of CPPVec can be found in Fig. 2, CPPVec mainly contains two steps, including feature extraction and classification model construction.

Fig. 2
figure 2

Pipeline of CPPVec. Multiple features are extracted from three kinds of sequence: the RNA sequence, the longest ORF extracted from the RNA sequence, and protein sequence translated from the longest ORF, and finally integrated into a SVM classifier for coding potential prediction. Note that the difference between CPPVec and CPPred lies in that the additional feature of doc2vec and the fixed feature of hexamer score

Full size image

During the process of feature extraction, the dataset (e.g., mRNAs, ncRNAs) is split into a training set and testing set, then the longest ORF, protein sequence are generated to calculate features, including four features from the longest ORF, four features from protein sequence and two features from RNA sequences. Note that the training and testing set are put together to generate distributed vector representations by doc2vec. Moreover, CPPVec calculated hexamer score with the first reading frame of the longest ORF instead of RNA sequence used in CPPred. We fixed this feature for the fact that the first reading frame of the longest ORF is likely to be the correct reading frame [14] and the calculation of hexamer score are more significant than that in the first reading frame of RNA sequence. As for classification model construction, we selected libsvm [38] as a classification model, the features of training and testing set were fed into the SVM classifier to train and test the classification model, respectively. Here, we choose SVM for the reasons: 1) Use the same classifier as CPPred to verify the effectiveness of additional features, e.g., doc2vec and fixed hexamer scores; 2) In CPPVec, the dimension of features is not high, and the scale of the datasets is not very large; 3) SVM has good implementation, e.g., libsvm is easy to use.

We chose the optimal parameters in CPPVec by gradually increasing one parameter with the other fixed, and observed the highest MCC scores achieved on Human-Training. In doc2vec, the context length was set to 4, the dimension of generated features was 100, and the “word” length was 3. As for SVM, the radial basis function was selected as the kernel function, the parameter C was set to 300 and gamma was 0.4. We also attempted to use grid.py script of libsvm for optimal C and gamma but it was very time-consuming. Moreover, the same setting as Human-Training is used for other datasets and we found that it worked well.

Performance of CPPVec on benchmark datasets

In order to verify the effectiveness of our proposed method, we compared our proposed method, CPPVec, with existing state-of-the-art methods, including CPPred, CPAT, CPC2, and PLEK. All the methods are trained and tested with the same datasets used in CPPred for a fair comparison. Human-Model is test on human, mouse, zebrafish, S. cerevisiae and fruit fly and Integrated-Model is test on Integrated-Testing.

From Tables 1, 2 and 3, it is observed that CPPVec performs the best among the existing state-of-the-art methods on all the test datasets. The MCCs of CPPVec are 0.953, 0.972 and 0.961 on Human-Testing, Mouse-Testing and Integrated-Testing, respectively, an improvement of 0.018 over the second best result achieved by PLEK on Human-Testing, 0.046 over the second best results achieved by CPPred on Mouse-Testing and 0.042 over the second best result achieved by CPPred on Integrated-Test, respectively. Moreover, we also test CPPVec on several other species to assess its performance on cross-species coding potential prediction. As shown in Tables 4, 5 and 6, CPPVec achieved consistent results when testing with zebrafish, S. cerevisiae and fruit fly, all of the AUC scores on the three testing sets exceed 0.99.

Table 1 Comparison of CPPVec (Human-Model), CPPred, CPAT, CPC2, and PLEK on Human-Testing
Full size table
Table 2 Comparison of CPPVec (Human-Model), CPPred, CPAT, CPC2, and PLEK on Mouse-Testing
Full size table
Table 3 Comparison of CPPVec (Human-Model), CPPred, CPAT, CPC2, and PLEK on Zebrafish-Testing
Full size table
Table 4 Comparison of CPPVec (Human-Model), CPPred, CPAT, CPC2, and PLEK on S.cerevisiae-Testing
Full size table
Table 5 Comparison of CPPVec (Human-Model), CPPred, CPAT, CPC2, and PLEK on Fruit-fly-Testing
Full size table
Table 6 Comparison of CPPVec (Integrated-Model), CPPred, CPAT, CPC2, and PLEK on Integrated-Testing
Full size table

Performance of CPPVec on experimentally validated lncRNAs

In order to verify the ability of CPPVec to identify novel lncRNAs, we compare the prediction performance of CPPVec with CPPred, CPAT, CPC2, and PLEK on 34 experimentally validated lncRNAs. As shown in Supplementary Table S1, CPPVec outperforms the other methods, only missing one lncRNAs with sequence name “NR_073054.1”. In comparison, CPPred, CPAT, CPC2 and PLEK incorrectly predict several other lncRNAs apart from “NR_073054.1”. Moreover, CPPVec correctly predicts lncRNAs with more confidence (smaller predicted coding probability) than the other methods, e.g., NR_111959.1, XR_593181.2, etc. All the results demonstrated that CPPVec has stronger ability to identify novel lncRNAs than existing state-of-the-art methods.

Performance of distributed representation

In order to verify the effectiveness of distributed representation of protein sequence translated from RNA sequence, we conducted an ablation study to separate the features used in CPPVec and observe the performance improvement that distributed feature vector contributes. To be specific, we use OVEC to denote the method that only use the 100 dimensional feature vector generated from doc2vec, we use NVEC to denote the method that use features of CPPVec except distributed features. All the methods are test on Integrated-Model using hold-out and 3-fold cross-validation. As shown in Supplementary Table S2, OVEC achieves MCC with 0.925, which even outperform CPPred that use multiple features. From Supplementary Table S3, OVEC achieves MCC with 0.912, which achieves considerable performance with CPPred. Moreover, we also analyzed the vector representations generated by doc2vec and k-mer on Integrated-Training dataset. Figure 3 shows a two-dimensional projection of generated vector representations by k-mer and doc2vec using t-SNE [39]. We can see that almost all the mRNAs and ncRNAs were clustered in two groups for doc2vec, whereas mRNAs mixed with ncRNAs heavily and difficult to separate for k-mer. It is interesting to see that there are two subgroups for ncRNAs features generated by doc2vec, this is because there exist multiple kinds of ncRNAs (e.g., piRNA, lncRNA, etc.) in Integrated-Training dataset and ncRNAs belong to the same class often show similar distributed representation. All the above results demonstrate that distributed representation of protein sequence is effective to distinguish mRNAs from ncRNAs.

Fig. 3
figure 3

Visualization of two-dimensional projections for k-mer and doc2vec using t-SNE

Full size image

Performance of fixed hexamer score

In order to verify the effectiveness of fixed hexamer score, we compared the prediction performance of NVEC and CPPred on Integrated-Test to observe the performance improvement of fixed hexamer score. From Supplementary Table S2, NVEC shows much better prediction performance than CPPred with MCC of 0.935 versus 0.919, and from Supplementary Table S3, NVEC shows much better prediction performance than CPPred with MCC of 0.937 versus 0.923, which verifies the significance of fixed hexamer score.

Conclusion

In this paper, we proposed a novel coding potential predictor (CPPVec) based on a distributed representation (e.g., doc2vec) of protein sequence translated from the longest ORF of RNA sequence, which effectively exploit the contextual information of protein sequence. Tests on human, mouse, fruit fly, zebrafish and S.cerevisiae demonstrates that CPPVec consistently outperforms existing state-of-the-art methods, which verifies that distributed representation of protein sequence is an effective feature for coding potential prediction.

Availability of data and materials

The source code of CPPVec are publicly available at: https://github.com/hgcwei/CPPVec. The datasets used in the paper are available at http://www.rnabinding.com/CPPred/. All other data that support the results of this study are available from the corresponding author upon request.

References

  1. Mercer TR, Dinger ME, Mattick JS. Long non-coding RNAs: insights into functions. Nat Rev Genet. 2009;10(3):155–9.

    Article  CAS  PubMed  Google Scholar 

  2. Wang X, Arai S, Song X, et al. Induced ncRNAs allosterically modify RNA-binding proteins in cis to inhibit transcription. Nature. 2008;454(7200):126.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  3. Wapinski O, Chang HY. Long noncoding rnas and human disease. Trends Cell Biol. 2011;21(6):354–61.

    Article  CAS  PubMed  Google Scholar 

  4. Schmitt AM, Chang HY. Gene regulation: Long RNAs wire up cancer growth. Nature. 2013;500(7464):536–7.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  5. Uchida S, Dimmeler S. Long Noncoding RNAs in Cardiovascular Diseases. Circ Res. 2013;116(4):737–50.

    Article  Google Scholar 

  6. Bao ZY, Yang Z, et al. LncRNADisease 2.0: an updated database of long non-coding RNA-associated diseases. Nucleic Acids Res. 2019;47(D1):D1034–7.

    Article  CAS  PubMed  Google Scholar 

  7. Ulitsky I, Bartel, et al. lincRNAs: Genomics, Evolution, and Mechanisms. Cell. 2013;154(1):26–46.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  8. Magny EG, Pueyo JI, Pearl F, et al. Conserved regulation of cardiac calcium uptake by peptides encoded in small open reading frames. Science. 2013;341(6150):1116–20.

    Article  CAS  PubMed  Google Scholar 

  9. Loman NJ, Misra RV, Dallman TJ, et al. Performance comparison of benchtop high-throughput sequencing platforms. Nat Biotechnol. 2013;30(5):434–9.

    Article  Google Scholar 

  10. Li A, Zhang J, Zhou Z. PLEK: a tool for predicting long non-coding RNAs and messenger RNAs based on an improved k-mer scheme. BMC Bioinformatics. 2014;15(1):1–10.

    Article  Google Scholar 

  11. Kang YJ, Yang DC, Kong L, et al. CPC2: a fast and accurate coding potential calculator based on sequence intrinsic features. Nucleic Acids Res. 2017;45(W1):W12–6.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  12. Wang L, Park HJ, Dasari S, et al. CPAT: Coding-Potential Assessment Tool using an alignment-free logistic regression model. Nucleic Acids Res. 2013;41(6):e74.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  13. Tong X, Liu S. CPPred: coding potential prediction based on the global description of RNA sequence. Nucleic Acids Res. 2019;47(8):e43.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  14. Furuno M, Kasukawa T, Saito R, et al. CDS annotation in full-length cDNA sequence. Genome Res. 2003;13(6B):1478–87.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  15. Meyer M, Stenzel U, Hofreiter M. Parallel tagged sequencing on the 454 platform. Nat Protoc. 2008;3(2):267–78.

    Article  CAS  PubMed  Google Scholar 

  16. Sun L, Luo H, Bu D, et al. Utilizing sequence intrinsic composition to classify protein-coding and long non-coding transcripts. Nucleic Acids Res. 2013;41(17):e166.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  17. Wen JH, Liu YS, et al. A classification model for lncRNA and mRNA based on k-mers and a convolutional neural network. BMC Bioinformatics. 2019;20(1):469.

    Article  PubMed  PubMed Central  Google Scholar 

  18. Ghandi M, Lee D, Mohammad-Noori M, et al. Enhanced regulatory sequence prediction using gapped k-mer features. PLoS Comput Biol. 2014;10(7):e1003711.

    Article  PubMed  PubMed Central  Google Scholar 

  19. Liu J, Gough J, Rost B. Distinguishing protein-coding from non-coding RNAs through support vector machines. PLoS Genet. 2006;2(4):e29.

    Article  PubMed  PubMed Central  Google Scholar 

  20. Dubchak I, Muchnik I, Holbrook SR, et al. Prediction of protein folding class using global description of amino acid sequence. Proc Natl Acad Sci. 2006;92(19):8700–4.

    Article  Google Scholar 

  21. Kong L, Zhang Y, et al. CPC: assess the protein-coding potential of transcripts using sequence features and support vector machine. Nucleic Acids Res. 2007;35(Web Server issue):W345–9.

    Article  PubMed  PubMed Central  Google Scholar 

  22. Lin MF, Jungreis I, et al. PhyloCSF: a comparative genomics method to distinguish protein coding and non-coding regions. Bioinformatics. 2011;27(13):i275–82.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  23. Bateman A, Martin MJ, et al. UniProt: a worldwide hub of protein knowledge. Bioinformatics. 2019;47(D1):D506–15.

    Google Scholar 

  24. Le Q, Mikolov T. Distributed representations of sentences and documents. Int Conf Mach Learn. 2014;1188–1196

  25. O’Leary NA, Wright MW, Brister JR, et al. Reference sequence (RefSeq) database at NCBI: current status, taxonomic expansion, and functional annotation. Nucleic Acids Res. 2016;44(D1):D733–45.

    Article  PubMed  Google Scholar 

  26. Zerbino DR, Achuthan P, Akanni W, et al. Ensembl 2018. Nucleic Acids Res. 2018;46(D1):D754–61.

    Article  CAS  PubMed  Google Scholar 

  27. Li W, Godzik A. Cd-hit: a fast program for clustering and comparing large sets of protein or nucleotide sequences. Bioinformatics. 2006;22:1658–9.

    Article  CAS  PubMed  Google Scholar 

  28. Zhou BL, Ji BH, et al. EVLncRNAs 2.0: an updated database of manually curated functional long non-coding RNAs validated by low-throughput experiments. Nucleic Acids Res. 2021;49(D1):D86–91.

    Article  CAS  PubMed  Google Scholar 

  29. Bengio Y, Courville A, Vincent P. Representation Learning: A Review and New Perspectives. IEEE Trans Patt Anal Mach Intell. 2013;35(8):1798–828.

    Article  Google Scholar 

  30. Bollegala D, Mu T, Goulermas JY. Cross-domain sentiment classification using sentiment sensitive embeddings. IEEE Trans Knowl Data Eng. 2015;28(2):398–410.

    Article  Google Scholar 

  31. Stein RA, Jaques PA, Valiati JF. An analysis of hierarchical text classification using word embeddings. Inf Sci. 2019;471:216–32.

    Article  Google Scholar 

  32. Kuriyan J, Konforti B, Wemmer D. The molecules of life: Physical and chemical principles. Garland Science; 2012

  33. Asgari E, Mofrad MRK. Continuous distributed representation of biological sequences for deep proteomics and genomics. PloS ONE. 2015;10(11):e0141287.

    Article  PubMed  PubMed Central  Google Scholar 

  34. Asgari E, Poerner N, McHardy A C, et al. DeepPrime2Sec: deep learning for protein secondary structure prediction from the primary sequences. bioRxiv. 2019;705426.

  35. Pan X, Shen HB. Learning distributed representations of RNA sequences and its application for predicting RNA-protein binding sites with a convolutional neural network. Neurocomputing. 2018;305:51–8.

    Article  Google Scholar 

  36. Deng L, Liu Y, Shi Y, et al. Deep neural networks for inferring binding sites of RNA-binding proteins by using distributed representations of RNA primary sequence and secondary structure. BMC Genomics. 2020;21(13):1–10.

    Google Scholar 

  37. Wei C, Zhang J, Yuan X. Enhancing the prediction of protein coding regions in biological sequence via a deep learning framework with hybrid encoding. Dig Signal Process. 2022;123:103430.

    Article  Google Scholar 

  38. Chang CC, Lin CJ. LIBSVM: a library for support vector machines. ACM Trans Intell Syst Technol. 2011;2(3):1–27.

    Article  Google Scholar 

  39. Maaten LVD, Hinton G. Visualizing data using t-SNE. Mach Learn Res. 2008;9(Nov):2579–605.

    Google Scholar 

Download references

Acknowledgements

The authors thank the anonymous reviewers for their valuable suggestions.

Funding

This work is supported by the Department of Science and Technology of Hubei Province under Grants 2020BAB012.

Author information

Authors and Affiliations

  1. School of Computer Science, Hubei University of Technology, Wuhan, China

    Chao Wei & Zhiwei Ye

  2. School of Computer Science and Technology, Xidian University, Xi’an, China

    Junying Zhang

  3. School of Computer Science and Engineering, Xi’an University of Technology, Xi’an, China

    Aimin Li

Authors
  1. Chao Wei
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Zhiwei Ye
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Junying Zhang
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Aimin Li
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

Conceptualization, Methodology, Data curation, Writing-original draft preparation, CW; Funding acquisition, ZWY; Writing-review and editing, JYZ, AML. The authors read and approved the final manuscript.

Corresponding author

Correspondence to Chao Wei.

Ethics declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Wei, C., Ye, Z., Zhang, J. et al. CPPVec: an accurate coding potential predictor based on a distributed representation of protein sequence. BMC Genomics 24, 264 (2023). https://doi.org/10.1186/s12864-023-09365-7

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s12864-023-09365-7

Keywords

BMC Genomics

ISSN: 1471-2164

Contact us