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

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate splice site identification is crucial for predicting gene structure.
• Machine learning approaches (MLAs) outperform rule-based methods but require numerical input via sequence encoding.
• Evaluating diverse encoding schemes is essential for optimizing MLA performance in splice site prediction.

Purpose of the Study:

• To comprehensively evaluate eight sequence encoding schemes for predicting donor and acceptor splice sites.
• To compare the performance of five supervised learning methods (ANN, Bagging, Boosting, RF, SVM) across different encoding schemes and species.
• To identify the optimal combination of encoding scheme and machine learning method for accurate splice site prediction.

Main Methods:

• Evaluated eight sequence encoding schemes (Bayes kernel, DS, DM, FDDM, FDTF, MM1, MM1+MM2, MM2).
• Utilized five supervised learning methods: Artificial Neural Network (ANN), Bagging, Boosting, Random Forest (RF), and Support Vector Machine (SVM).
• Tested and validated performance across multiple species, including A. thaliana, C. elegans, D. melanogaster, H. sapiens, Ciona intestinalis, Dictyostelium discoideum, Phaeodactylum tricornutum, and Trypanosoma brucei.

Main Results:

• The paired-nucleotide frequency difference (FDTF) encoding scheme achieved the highest accuracy, followed by MM2 or FDDM.
• Support Vector Machine (SVM) and Random Forest (RF) demonstrated superior performance among the tested machine learning methods.
• The SVM-FDTF combination proved to be the optimal approach for splice site prediction, with higher accuracy observed in species with lower intron density.

Conclusions:

• The FDTF encoding scheme combined with SVM offers a highly accurate method for splice site prediction.
• This study provides a comprehensive benchmark for sequence encoding strategies in splice site identification.
• An R-package, EncDNA, has been developed to facilitate the application of these encoding schemes in genomic analysis.