Performing Cancer Diagnosis via an Isoform Expression Ranking-based LSTM Model

Abstract

The known set of genetic factors involved in the development of several types of cancer has considerably been expanded, thus easing to devise and implement better therapeutic strategies. The automatic diagnosis of cancer, however, remains as a complex task because of the high heterogeneity of tumors and the biological variability between samples. In this work, a long short-term memory network-based model is proposed for diagnosing cancer from transcript-base data. An efficient method that transforms data into gene/isoform expression-based rankings was formulated, allowing us to directly embed important information in the relative order of the elements of a ranking that can subsequently ease the classification of samples. The proposed predictive model leverages the power of deep recurrent neural networks, being able to learn existing patterns on the individual rankings of isoforms describing each sample of the population. To evaluate the suitability of the proposal, an extensive experimental study was conducted on 17 transcript-based datasets, and the results showed the effectiveness of this novel approach and also indicated the gene/isoforms expression-based rankings contained valuable information that can lead to a more effective cancer diagnosis.