A transfer-learning approach to feature extraction from cancer transcriptomes with deep autoencoders

Abstract

The diagnosis and prognosis of cancer are among the more challenging tasks that oncology medicine deals with. With the main aim of fitting the more appropriate treatments, current personalized medicine focuses on using data from heterogeneous sources to estimate the evolu- tion of a given disease for the particular case of a certain patient. In recent years, next-generation sequencing data have boosted cancer prediction by supplying gene-expression information that has allowed diverse machine learning algorithms to supply valuable solutions to the problem of cancer subtype classification, which has surely contributed to better estimation of patient’s response to diverse treatments. However, the efficacy of these models is seriously affected by the existing imbalance between the high dimensionality of the gene expression feature sets and the number of sam- ples available for a particular cancer type. To counteract what is known as the curse of dimensionality, feature selection and extraction methods have been traditionally applied to reduce the number of input variables present in gene expression datasets. Although these techniques work by scaling down the input feature space, the prediction performance of tradi- tional machine learning pipelines using these feature reduction strategies remains moderate. In this work, we propose the use of the Pan-Cancer dataset to pre-train deep autoencoder architectures on a subset com- posed of thousands of gene expression samples of very diverse tumor types. The resulting architectures are subsequently fine-tuned on a col- lection of specific breast cancer samples. This transfer-learning approach aims at combining supervised and unsupervised deep learning models with traditional machine learning classification algorithms to tackle the problem of breast tumor intrinsic-subtype classification.