Generative deep learning for biomedical data analysis

Abstract

In my article, the author selected two types of breast cancer samples, Ductal carcinoma in situ(DCIS) and Lobular Carcinoma. The authors selected 35 Ductal carcinoma in situ samples and 35 Lobular Carcinoma samples from the TCGA database. After non-specific filtering of the samples for low expression genes in the RNAseq data, only 636 genes were left in each group. The retained genes were used for expression differential analysis studies. The author used heat maps and principal component analysis to visualize the actual impact of each gene. Afterward, a Variational Auto-encoder model was built to simulate the generation of new gene sequences. The specific model trained in this study consists of gene expression input (the 636 most variably expressed genes by median absolute deviation) compressed into two vectors of length 100 (mean and variance coding space), which are made deterministic by a re-parameterization technique that draws ε-vectors from the uniform distribution. The coding layer is then decoded back to the original 636 dimensions by a single reconstruction layer. The encoding scheme also uses relu activation, while the decoder uses sigmoid activation to perform forward activation. All weights are initialized uniformly by Glorot. Finally, we can see that the values of the generated sequences are relatively close to the original sequences.