Conditional generative adversarial network for gene expression inference

[stephenra]

The rapid progress of gene expression profiling has facilitated the prosperity of recent biological studies in various fields, where gene expression data characterizes various cell conditions and regulatory mechanisms under different experimental circumstances. Despite the widespread application of gene expression profiling and advances in high-throughput technologies, profiling in genome-wide level is still expensive and difficult. Previous studies found that high correlation exists in the expression pattern of different genes, such that a small subset of genes can be informative to approximately describe the entire transcriptome. In the Library of Integrated Network-based Cell-Signature program, a set of ∼1000 landmark genes have been identified that contain ∼80% information of the whole genome and can be used to predict the expression of remaining genes. For a cost-effective profiling strategy, traditional methods measure the profiles of landmark genes and then infer the expression of other target genes via linear models. However, linear models do not have the capacity to capture the non-linear associations in gene regulatory networks.

https://doi.org/10.1093/bioinformatics/bty563

[gwaybio]

Wang et al. apply a conditional generative adversarial network (GGAN; not sure what the first G stands for) to the LINCS problem of gene expression inference. The authors compare GGAN to three linear methods, a k nearest neighbor approach, and D-GEX. We discuss D-GEX in #24. Tagging @dhimmel because of his previous interest/experience in LINCS and his comments on the previous issue.

Summary

• GAN conditioned on predicting corresponding target genes from a given gene expression profile's landmark genes.
• The discriminator loss is a least squares difference between generated predicted genes and actual predicted genes.
  • Both generated and predicted genes concatenate the same target gene values
  • 1 layer - Fully connected with 3000 hidden units
• Generator is a DenseNet which is convolutional
  • 3 layers with 9000 hidden units each
  • It is unclear how DenseNet is used without spatial structure - the authors discuss masking input into the discriminator and slowing increasing the amount of exposed genes, but I am not clear how this would alleviate the lack of spatial structure.
• Evaluation of predicting 10% holdout GEO and cross platform prediction of GTEx
  • Because the paper mirrors the experimental setup of the D-GEX paper, the GGAN authors lift the results from the D-GEX paper for their comparison
  • GGAN and D-GEX outperform linear and knn methods in mean absolute error and concordance correlation
  • GGAN outperforms D-GEX
• For biological interpretation, the authors map clusters of landmark genes to their contribution to predicted target genes using layer-wise relevance propagation.

[stephenra]

@gwaygenomics It isn't explicitly mentioned anywhere but I believe the 'G' is for 'gene expression'.