You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
When looking into the model weights of some of the predicted TFs (20200223.trained) there were some weird things i noticed:
some TFs have a negative motif weight, with the most extreme examples being CDC5L, IRF5, NFATC1, NEUROD1 and PAX8. This is odd, if the absence of a motif is indication that a TF binds there in the chipseq experiments, shouldn't the motif be discarded?
some TFs have a really really nonspecific motif weight/score but also really high distance paramter E.g. SREBF2, apparently (or at least how i interpret that) how ananse predicts that TF binding is: "Is there a close ATAC peak to the TSS? SREBF2 might regulate that". Which I think is also a bit eeh finicky when it comes to follow up. Especially problematic when other TFs are matched to the model trained on SREBF2 data, even though that model seems to imply the motif the TFs are matched on isnt really contributing anything in the first place.
I'm not sure if i interpret those things correctly, and perhaps they are also not things to easily fix but thought I'll type outa issue.
The text was updated successfully, but these errors were encountered:
When looking into the model weights of some of the predicted TFs (20200223.trained) there were some weird things i noticed:
some TFs have a negative motif weight, with the most extreme examples being CDC5L, IRF5, NFATC1, NEUROD1 and PAX8. This is odd, if the absence of a motif is indication that a TF binds there in the chipseq experiments, shouldn't the motif be discarded?
some TFs have a really really nonspecific motif weight/score but also really high distance paramter E.g. SREBF2, apparently (or at least how i interpret that) how ananse predicts that TF binding is: "Is there a close ATAC peak to the TSS? SREBF2 might regulate that". Which I think is also a bit eeh finicky when it comes to follow up. Especially problematic when other TFs are matched to the model trained on SREBF2 data, even though that model seems to imply the motif the TFs are matched on isnt really contributing anything in the first place.
I'm not sure if i interpret those things correctly, and perhaps they are also not things to easily fix but thought I'll type outa issue.
The text was updated successfully, but these errors were encountered: