Massive Production of Cancer Synthetic RNA-Seq Gene Expression Samples

Abstract

The financial and practical complexities of collecting a statistically significant number of tumour gene expression data canseverely affect the reliability of differential expression analysis in cancer. Generally, studies are performed with an immensenumber of genes related to a minimal number of samples, leading to an underrepresentation of the disease heterogeneity. Apotential solution lies in generating artificial tumour samples similar to real ones, thereby increasing the accuracy and robust-ness of biological experiments. In that sense, this work proposes a Wasserstein Generative Adversarial Network architecturefor generating synthetic tumour samples. The primary objective is to ensure that the generated samples accurately preservethe shared behaviour of genes and the differential expression patterns observed in tumour samples. The study focuses onfour cancer types: thyroid, breast, lung and prostate. The experimental results confirm that our proposal can reproduce thecharacteristics of the original samples, suggesting that the model is a reliable framework for increasing the number of tumoursamples. In addition, the model proved able to exacerbate the distributions shown by each gene, allowing the discovery ofpossible hidden patterns.