Learning a hierarchical representation of the yeast transcriptomic machinery using an autoencoder model

[cgreene]

https://dx.doi.org/10.1186/s12859-015-0852-1

[agitter]

The precise biological meaning ascribed to each of the hidden layers in Figure 1c may risk overinterpreting the model.

[cgreene]

@agitter : Figure 3 is surprisingly clean for the first layer. I agree that the evidence for the later levels in "The hierarchical structure captures signals of different degrees of abstraction" seems a bit much for the amount of evidence. Probably good to discuss alongside #6 #22 & the rest in a nucleus of gene expression + unsupervised feature construction papers.

[cgreene]

Include with unsupervised feature construction. May also be a useful example for of potential dangers of too much interpretation cc @akundaje

[gwaybio]

Just revisited this paper - I'm going to reopen it to cite in the gene expression study section (It doesn't appear that we've stuck to the "closing an issue when its not being discussed" paradigm). I agree with the over interpretation concerns, but the concept is very cool. The paper also had a nice discussion of limitations of current approaches.

Biological Aspects

• Yeast cDNA microarrays (1609 from puma.princeton.edu and 300 from Hughes et al 2000)
• Found 2228 genes to be differentially expressed between control and treatments (variety of treatments; chemical, gene mutation, etc.)
• Binary entries 0 or 1 if gene was differentially expressed or not

Computational Aspects

• Compared NMF vs. Autoencoder vs. Sparse Autoencoder
• 4 hidden layers for autoencoders
• Evaluation based on reconstruction error and BIC
• Also evaluated how well Autoencoder learned features performed in a clustering analysis compared to raw and NMF features (Fig 5)