DeepCancer: Detecting Cancer through Gene Expressions via Deep Generative Learning

[sw1]

Transcriptional profiling on microarrays to obtain gene expressions has been used to facilitate cancer diagnosis. We propose a deep generative machine learning architecture (called DeepCancer) that learn features from unlabeled microarray data. These models have been used in conjunction with conventional classifiers that perform classification of the tissue samples as either being cancerous or non-cancerous. The proposed model has been tested on two different clinical datasets. The evaluation demonstrates that DeepCancer model achieves a very high precision score, while significantly controlling the false positive and false negative scores.

https://arxiv.org/abs/1612.03211

They use a GAN and compare it to an RBM.

[gwaybio]

Computational Methods

Generative Adversarial Network

• Generator - CNN that takes 100 random vector (z) input and outputs 183 x 224 matrix
• Discriminator - CNN that takes in Generated and Real 183 x 224 matrix of gene expression values

(183 x 224 = 40,992 probes) - they did not discuss how the probes were ordered.

It was interesting that their motivation for using a GAN was to use the probe weights that the Discriminator learned to classify cancer vs. non cancer. It seems to me that the weights would learn real vs. fake and not necessarily have any direct bearing on cancer vs. non cancer. Makes me wonder why a GAN is used here at all - they do try to compare with SVM and logistic regression but results are not very clear.

Biological Relevance

• Predicting if a sample was normal, inflammatory or cancerous breast (GSE45584)
• Predicting if prostate gene expression data was normal or cancerous

Other comments

• No discussion of potential batch effect correction
• Name of their algorithm changes from DeepCancer to DeepCloud by the end of the paper
• I think their discussion of training the GAN is valuable and I believe that with this architecture they can generate pretty realistic samples
  • There is no actual proof that they are generating well
• I would say that the domain is really cool but a lot more can be done with GANs + Gene Expression

[agitter]

@gwaygenomics Do you think we should include this in the review, and if so, does it belong in Study (gene expression sub-section) or Categorize?

[gwaybio]

Agree with categorize label. Not sure about inclusion in review. Generative model plus categorize is really interesting and I think could have a big impact but this article doesn't really use the generator to its potential. I guess I'd say let's see how the categorize section comes together

[agitter]

We also have the potential to stick something like that in the Discussion. I have been thinking about network architectures that are hot in other domains but not yet widely applied in biomedicine. Those may be areas of opportunity in the short term, and GANs could fit in that category.

[cgreene]

I'm going to drop this for now from consideration for the categorize section. I think that, given the issues of the paper, it's not quite ready to help us determine whether or not this pushes the field forward.