search

griffithlab / DeepSVR

MIT License
50 stars 19 forks source link

All candidates are predicted as true mutations #16

Open ozlemmuslu opened 4 years ago

ozlemmuslu commented 4 years ago

Dear DeepSVR team,

Thank you for providing this tool for the community.

I have tried it on two cell lines with known mutation sites (majority are true mutations) and realized that the tool predicts every site to be a true mutation, and gives the exact same score (for ambiguous, fail, and somatic) for each of them. I first used your data to train a model, when that didn't work, I used other cell lines to train one, with similar results.

I also tweaked the number of epochs, and I got slightly different scores for different sites, but still everything was predicted as true mutations.

I'm wondering if you have an already trained model that I can directly use. I'm especially interested in your TCGA model (I have TCGA access), and whether that is available/could be made available under controlled access. If not, do you have any suggestions for training the model?

Thanks!

zlskidmore commented 4 years ago

Hi @ozlemmuslu

We have noticed a couple issues/bugs that have cropped up with DeepSVR and are attempting to find the bandwith to correct these. The underlying Data is still available in the /Data folder.