Feature selection may improve deep neural networks for the bioinformatics problems.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• Deep neural networks (DNNs) show promise for predicting biomedical phenotypes.
• Feature selection has not been extensively explored to optimize DNN performance.

Purpose of the Study:

• To investigate whether feature selection algorithms can improve DNN predictive accuracy.
• To compare the effectiveness of various feature selection methods on different DNN architectures and datasets.

Main Methods:

• Evaluated 11 feature selection algorithms across 6 DNN models (CNN, DBN, RNN, MobilenetV2, ShufflenetV2, Squeezenet).
• Tested on 5 methylomic and 19 transcriptome datasets (binary and multi-class classification).
• Implemented and tested algorithms in Python 3.6.6.

Main Results:

• Feature selection algorithms generally improved DNN model performance.
• Deep belief networks (DBNs) combined with SVM-RFE achieved top prediction accuracies on methylomic datasets.
• The positive impact of feature selection was observed across different data types.

Conclusions:

• Feature selection is a valuable strategy for enhancing DNN-based biomedical predictions.
• The choice of feature selection algorithm and DNN architecture impacts performance.
• DBN models with SVM-RFE offer a robust approach for methylomic data analysis.