Top-k calculations about spliceai HOT 6 CLOSED

I'm not sure I understood the "averaged top-k accuracy" phrase. Are you calculating the top-k accuracy of each of the five models one by one? If so, the 92-93% is along the lines of what we observed. You are supposed to average the output of the model (ensemble prediction) and then calculate the top-k accuracy. Let me know if that is what you are doing (i.e. running the code without changing it).

from spliceai.

Yes, those results are from running the code (which I didn't change). I meant average across both the acceptor and donor classes. If those numbers seem right, then how is the number in the paper (95% top-k accuracy) calculated? Thanks!

from spliceai.

Where did you get the models from? If it is from the github repo, then that explains your results - those models are trained using the GTEx folder and not the Canonical folder. If you train models using the files from the Canonical folder, you should see the 95% reported in the paper.

from spliceai.

Got it, are the Canonical models available online?

from spliceai.

Could you try this link: https://drive.google.com/file/d/1DrnOVmyLV2rFWWTa-lbZWzP3YzwO59K2/view (the models with _c*.h5 in their names) are what you are interested in.

from spliceai.

Thanks!

from spliceai.