An Improved Method for Splice Site Prediction in DNA Sequences Using Support Vector Machines

From shallow to deep: some lessons learned from application of machine learning for recognition of functional genomic elements in human genome

Identification of genomic signals as indicators for functional genomic elements is one of the areas that received early and widespread application of machine learning methods. With time, the methods applied grew in variety and generally exhibited a tendency to improve their ability to identify some major genomic and transcriptomics signals. The evolution of machine learning in genomics followed a similar path to applications of machine learning in other fields. These were impacted in a major way by three dominant developments, namely an enormous increase in availability and quality of data, a significant increase in computational power available to machine learning applications, and finally, new machine learning paradigms, of which deep learning is the most well-known example. It is not easy in general to distinguish factors leading to improvements in results of applications of machine learning. This is even more so in the field of genomics, where the advent of next-generation sequencing and the increased ability to perform functional analysis of raw data have had a major effect on the applicability of machine learning in OMICS fields. In this paper, we survey the results from a subset of published work in application of machine learning in the recognition of genomic signals and regions in human genome and summarize some lessons learnt from this endeavor. There is no doubt that a significant progress has been made both in terms of accuracy and reliability of models. Questions remain however whether the progress has been sufficient and what these developments bring to the field of genomics in general and human genomics in particular. Improving usability, interpretability and accuracy of models remains an important open challenge for current and future research in application of machine learning and more generally of artificial intelligence methods in genomics.

Novel Transformer Networks for Improved Sequence Labeling in genomics

In genomics, a wide range of machine learning methodologies have been investigated to annotate biological sequences for positions of interest such as transcription start sites, translation initiation sites, methylation sites, splice sites and promoter start sites. In recent years, this area has been dominated by convolutional neural networks, which typically outperform previously-designed methods as a result of automated scanning for influential sequence motifs. However, those architectures do not allow for the efficient processing of the full genomic sequence. As an improvement, we introduce transformer architectures for whole genome sequence labeling tasks. We show that these architectures, recently introduced for natural language processing, are better suited for processing and annotating long DNA sequences. We apply existing networks and introduce an optimized method for the calculation of attention from input nucleotides. To demonstrate this, we evaluate our architecture on several sequence labeling tasks, and find it to achieve state-of-the-art performances when comparing it to specialized models for the annotation of transcription start sites, translation initiation sites and 4mC methylation in E. coli.

Splice2Deep: An ensemble of deep convolutional neural networks for improved splice site prediction in genomic DNA

Background

The accurate identification of the exon/intron boundaries is critical for the correct annotation of genes with multiple exons. Donor and acceptor splice sites (SS) demarcate these boundaries. Therefore, deriving accurate computational models to predict the SS are useful for functional annotation of genes and genomes, and for finding alternative SS associated with different diseases. Although various models have been proposed for the in silico prediction of SS, improving their accuracy is required for reliable annotation. Moreover, models are often derived and tested using the same genome, providing no evidence of broad application, i.e. to other poorly studied genomes.

Results

With this in mind, we developed the Splice2Deep models for SS detection. Each model is an ensemble of deep convolutional neural networks. We evaluated the performance of the models based on the ability to detect SS in Homo sapiens, Oryza sativa japonica, Arabidopsis thaliana, Drosophila melanogaster, and Caenorhabditis elegans. Results demonstrate that the models efficiently detect SS in other organisms not considered during the training of the models. Compared to the state-of-the-art tools, Splice2Deep models achieved significantly reduced average error rates of 41.97% and 28.51% for acceptor and donor SS, respectively. Moreover, the Splice2Deep cross-organism validation demonstrates that models correctly identify conserved genomic elements enabling annotation of SS in new genomes by choosing the taxonomically closest model.

Conclusions

The results of our study demonstrated that Splice2Deep both achieved a considerably reduced error rate compared to other state-of-the-art models and the ability to accurately recognize SS in other organisms for which the model was not trained, enabling annotation of poorly studied or newly sequenced genomes. Splice2Deep models are implemented in Python using Keras API; the models and the data are available at https://github.com/SomayahAlbaradei/Splice_Deep.git.

Evaluating the performance of sequence encoding schemes and machine learning methods for splice sites recognition

Identification of splice sites is imperative for prediction of gene structure. Machine learning-based approaches (MLAs) have been reported to be more successful than the rule-based methods for identification of splice sites. However, the strings of alphabets should be transformed into numeric features through sequence encoding before using them as input in MLAs. In this study, we evaluated the performances of 8 different sequence encoding schemes i.e., Bayes kernel, density and sparse (DS), distribution of tri-nucleotide and 1st order Markov model (DM), frequency difference distance measure (FDDM), paired-nucleotide frequency difference between true and false sites (FDTF), 1st order Markov model (MM1), combination of both 1st and 2nd order Markov model (MM1 + MM2) and 2nd order Markov model (MM2) in respect of predicting donor and acceptor splice sites using 5 supervised learning methods (ANN, Bagging, Boosting, RF and SVM). The encoding schemes and machine learning methods were first evaluated in 4 species i.e., A. thaliana, C. elegans, D. melanogaster and H. sapiens, and then performances were validated with another four species i.e., Ciona intestinalis, Dictyostelium discoideum, Phaeodactylum tricornutum and Trypanosoma brucei. In terms of ROC (receiver-operating-characteristics) and PR (precision-recall) curves, FDTF encoding approach achieved higher accuracy followed by either MM2 or FDDM. Further, SVM was found to achieve higher accuracy (in terms of ROC and PR curves) followed by RF across encoding schemes and species. In terms of prediction accuracy across species, the SVM-FDTF combination was optimum than other combinations of classifiers and encoding schemes. Further, splice site prediction accuracies were observed higher for the species with low intron density. To our limited knowledge, this is the first attempt as far as comprehensive evaluation of sequence encoding schemes for prediction of splice sites is concerned. We have also developed an R-package EncDNA (https://cran.r-project.org/web/packages/EncDNA/index.html) for encoding of splice site motifs with different encoding schemes, which is expected to supplement the existing nucleotide sequence encoding approaches. This study is believed to be useful for the computational biologists for predicting different functional elements on the genomic DNA.

Markovian encoding models in human splice site recognition using SVM

Splice site recognition is among the most significant and challenging tasks in bioinformatics due to its key role in gene annotation. Effective prediction of splice site requires nucleotide encoding methods that reveal the characteristics of DNA sequences to provide appropriate features to serve as input of machine learning classifiers. Markovian models are the most influential encoding methods that highly used for pattern recognition in biological data. However, a direct performance comparison of these methods in splice site domain has not been assessed yet. This study compares various Markovian encoding models for splice site prediction utilizing support vector machine, as the most outstanding learning method in the domain, and conducts a new precise evaluation of Markovian approaches that corrects this limitation. Moreover, a novel sequence encoding approach based on third order Markov model (MM3) is proposed. The experimental results show that the proposed method, namely MM3-SVM, performs significantly better than thirteen best known state-of-the-art algorithms, while tested on HS3D dataset considering several performance criteria. Further, it achieved higher prediction accuracy than several well-known tools like NNsplice, MEM, MM1, WMM, and GeneID, using an independent test set of 50 genes. We also developed MMSVM, a web tool to predict splice sites in any human sequence using the proposed approach. The MMSVM web server can be assessed at https://pashaei.shinyapps.io/mmsvm.