search

kipoi / models

Model zoo for genomics
http://kipoi.org
MIT License
165 stars 59 forks source link

Fix APARENT/site_probabilities output labels #285

Closed twrightsman closed 2 years ago

twrightsman commented 2 years ago

From the APARENT paper methods section on the cleavage model:

The 186 first Softmax probabilities are trained to predict the cleavage proportions of the 186 nucleotides in the input sequence. The 187-th probability predicts the remaining proportion of polyadenylation outside the sequence window (i.e., the distal isoform).

The model used by Kipoi is an updated one with 205bp but should be the same in terms of output arrangement.

I'm not sure if the tests also have to be updated, but I don't know how to do that.

haimasree commented 2 years ago

Hello! Thanks for the update and it seems tests are fine as well. @Hoeze I am merging this pr. Let me know if you disagree.

Hoeze commented 2 years ago

Hi @twrightsman, thanks a lot for your pull request, you're absolutely right! It's really cool that you found + fixed that mistake :grinning:

The same issue also affects the variant effect prediction. Also, I'm investigating how to best handle the isoform window there. PR will follow :)

Sorry for my late answer, I did not have time to investigate your PR earlier :no_mouth: