Hi, Deep learning people
I am new to VAE and wonder if the following framework fits with VAE in pytorch. As I am working on bioinformatics, let me explain the high-level idea based on biology.
Let’s say I have 10 mRNA isoforms (or you can just call it isoforms). These 10 isoforms came from the same gene as one gene can contain multiple isoforms due to alternative splicing. Anyway, I want to see which of these 10 isoforms is the most probable based on the 10 features that I have per isoform. For now let’s just say there are 10 features, so every isoform has 10 features. At the moment, I don’t know which one of those 10 features is important. Is it possible to train those 10 features of each isoform (so in this case there will be 10 * 10 = 100) and find which isoform is the most probable based on those 10 features? If this is possible, can you rank those 10 isoforms based on the latent space or reconstruction error in the decoder that learned the distribution of 10 features?
Or instead, is it common to just use the latent space of the VAE and feed that as an input to a feed-forward neural network to rank those 10 isoforms? In this case, let’s assume we have true labels.