Transforming OMIC features for classification using siamese convolutional networks.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Modern biotechnology generates vast amounts of omic data, including transcriptomes and methylomes.
• The
• large p, small n

Purpose of the Study:

• To address the classification challenges in omic data with a high feature-to-sample ratio.
• To develop and evaluate a novel feature engineering algorithm, SiaCo, for improved omic data classification.

Main Methods:

• A Siamese convolutional network was employed to transform omic features into a new feature space.
• The SiaCo algorithm was designed to minimize intra-class distances and maximize inter-class distances.
• The algorithm was evaluated on both transcriptome and methylome datasets.

Main Results:

• SiaCo features demonstrated improved classification accuracies for binary classification tasks on independent test datasets.
• The algorithm enhanced overall classification performance rather than individual feature discrimination power.
• SiaCo features, while effective, lack interpretability due to the network's transformation nature.

Conclusions:

• SiaCo offers a promising approach for enhancing classification in high-dimensional omic datasets.
• The method effectively improves predictive accuracy in transcriptomic and methylomic analyses.
• While interpretability is limited, SiaCo provides a valuable tool for omic data classification.