AvantiShri
commented
7 years ago If I recall correctly, entire-gene holdout should be added to (3).
For (5), I think to make it more like the square loss, you should square the magnitudes. BTW, If you want to avoid writing a new loss function, then one thing you can do is just apply the standard square loss to (1 - P(0))R but modify the labels such that the "true" value is 0 for non-differentially-regulated genes and the original value for differentially-regulated genes.
boxiangliu
To summarize our discussion from yesterday, here are the TODO items: 1) use a 2 convolutional layers. 2) stratify magnitude and variance to sample held-out genes. 3) hold out i) random and ii) entire series of stress conditions and iii) entire genes. 4) use i) dinucleotide shuffle and ii) whole-promoter shuffle to test whether trained model has learned regulatory motifs. 5) use zero-inflated loss. One form could be:||P(1)&space;R||^2&space;+&space;L_C&space;||P(1)&space;R-L_R||^2 "(1-L_C)||P(1) R||^2 + L_C ||P(1) R-L_R||^2")
where L_C is classification label and L_R is regression label. When L_C is 0 (baseline), the second term goes to 0, and we want P(1) to be large. When L_C is 1 (up- or down-regulated), the first term goes to 0 and we want P(0) to be close to 0. The network would have three outputs, each representing baseline, up/down-regulated, and regression output.
6) reduce the number of filter to ~50 on the first layer to avoid memorization.
7) augment the promoter set with dinucleotide shuffle to force motif learning.
8) use delta deepLIFT to detect motif-motif and motif-gene interactions.
9) Use Gorrila GO to compare predicted and actual pathways.
Feel free to add/correct items by commenting below.