differential gene expression analysis

Question

[maryak]

Can we use convolutional neural network for differential gene expression(RNA seq) analysis ? Any help in this regard will be appreciated

[maryak]

Actually i want to identify cancer related genes using RNA seq gene expression dataset

[russhh]

convolutional networks are for datasets where there is a sense of neighbourhood: for example, in a matrix of pixel intensities from an image you know that the position m[1, 1] is a neighbour of m[1, 2] and so you can use information from the neighbourhood of a given point to help define larger-scale features for use in subsequent levels of your neural network.

How do you see that translating over to RNA-Seq. To my mind there's three main sources of neighbourhood information pertinent to RNA-Seq: you could convolve the coverage at different bases within a transcript; or you could convolve over the different exons for a transcript; or you could use epigenetic marks in the neighbourhood of a gene to predict the response of that gene.

I haven't seen that sort of stuff in the wild; but there's loads of possibilities. Too busy doing the do-able myself

Answer 1

[Kevin Blighe]

The term 'convolutional neural network' does not fit that well with 'differential gene expression analysis'. What is it that you are trying to do?

A possible workflow is:

1. Normalise RNA-seq raw counts
2. Perform a differential expression analysis on the normalised counts to identify statistically significantly differentially expressed genes between conditions / treatments of interest.
3. Transform normalised counts to log or Z scale
4. Perform hierarchical clustering and generate a heatmap using the transformed, normalised counts, filtered for your genes that pass your chosen thresholds for statistical significance

Independently, one may do this:

1. Normalise RNA-seq raw counts
2. Transform normalised counts to log or Z scale
3. Create a neural network using your transformed, normalised counts in an attempt to see which genes are more associated with others by their inclusion in the same module - this is unbiased network construction

Note that some create networks from a pre-selected list of genes, such as genes that are statistically significantly differentially expressed. This would be biased network construction. You could also create separate networks for your control and disease groups, and then compare these networks to see how different genes associate with each other in the 2 conditions.

Kevin

[maryak]

"Transform normalised counts to log or Z scale" why do we do so ? what information does z score give?

[Kevin Blighe]

It tells you which genes are 1, 2, or 3 standard deviations from the mean expression value. Log2 values do not represent this.