Extracting a biologically relevant latent space from cancer transcriptomes with variational autoencoders

Abstract

The Cancer Genome Atlas (TCGA) has profiled over 10,000 tumors across 33 different cancertypes for many genomic features, including gene expression levels. Gene expression measurements capture substantial information about the state of each tumor. Certain classes of deep neural network models are capable of learning a meaningful latent space. Such a latent space could be used to explore and generate hypothetical gene expression profiles under various types of molecular and genetic perturbation. For example, one might wish to use such a model to predict a tumor's response to specific therapies or to characterize complex gene expression activations existing in differential proportions in different tumors. Variational autoencoders (VAEs) are a deep neural network approach capable of generating meaningful latent spaces for image and text data. In this work, we sought to determine the extent to which a VAE can be trained to model cancer gene expression, and whether or not such a VAE would capture biologically-relevant features. In the following report, we introduce a VAE trained on TCGA pan-cancer RNA-seq data, identify specific patterns in the VAE encoded features, and discuss potential merits of the approach. We name our method "Tybalt" after an instigative, cat-like character who sets a cascading chain of events in motion in Shakespeare's "Romeo and Juliet". From a systems biology perspective, Tybalt could one day aid in cancer stratification or predict specific activated expression patterns that would result from genetic changes or treatment effects.

Variational autoencoders for cancer data integration: design principles and computational practice

“…Autoencoders have been used for learning representations and analysing transcriptomic cancer data before. In particular, our work relates to Way and Greene (2018), since it employs VAEs for constructing latent representations and analysing transcriptomic cancer data. The authors show that VAEs can be utilised for knowledge extraction from gene expression pan-cancer TCGA data (TCGA et al, 2013), thus reducing the dimensionality of the single, homogeneous data source while still being able to identify patterns related to different cancer types.…”

Deep Genomic Signature for early metastasis prediction in prostate cancer

“…3) After training the second DAE, genes with high weights in this DAE are selected based on a standard deviation filter on their connectivity weights. This selection approach is similar to (Danaee et al, 2016;Tan et al, 2015Tan et al, , 2016Tan et al, , 2017Way and Greene, 2018). The idea of this gene selection step is to provide the classifier with a few rich genes with the strongest signals based on both the labelled and unlabelled datasets.…”

“…In the previous studies the value of this threshold for similar purposes was typically set to 2 (Danaee et al, 2016;Tan et al, 2015Tan et al, , 2016 or 2.5 (Tan et al, 2017;Way and Greene, 2018).…”

“…VAE and AAE were also successfully utilized in designing new molecules with desired properties for drug discovery purposes (Gómez-Bombarelli et al, 2016;Kadurin et al, 2017). Moreover, VAE was also able to capture patterns in the gene expression pan-cancer data for specific tissues (Way and Greene, 2018).…”

Applications of Genetics, Genomics and Bioinformatics in Drug Discovery

“…While deep learning has been applied in many domains such as speech recognition, image recognition, natural language processing, its application in analyzing genomic data is very limited. Way et al 17 applied variational autoencoders (VAEs), an unsupervised deep neural network approach to analyze TCGA gene expression profiling data. Specifically, the extent to which a VAE can be trained to model cancer gene expression, and whether or not such a VAE would capture biologically relevant features were evaluated.…”