jlistgarten
commented
6 years ago Hi Vineet,
Although the training data were in the range 0.0 to 1.0, the final trained regression model we use can make predictions outside of this range. Still, we expect these to be somewhat rare. So no, there is no bug, and what you describe sounds reasonable. If it’s easier for your purposes, you could just set values to the closest part of the range 0.0-1.0 (i.e. negative values to 0, and values greater than 1.1 to 1.0).
Jennifer
From: Vineet Gopal [mailto:notifications@github.com] Sent: Wednesday, August 9, 2017 6:56 PM To: MicrosoftResearch/Azimuth Azimuth@noreply.github.com Cc: Subscribed subscribed@noreply.github.com Subject: [MicrosoftResearch/Azimuth] Negative guide scores (#21)
Hello! We've noticed that occasionally, guides that we score will get a negative predicted score from azimuth. Here's an example:
from azimuth.model_comparison import predict
import numpy
predict(numpy.array(['AACTGATTTCTGGCGTTTTCTTTCTGGCTC']), numpy.array([8905]), numpy.array([96]))
No model file specified, using V3_model_full
array([-0.04603427])
Empirically, it looks like negative scores are more likely to happen with peptide percentages closer to 100. However, from the Azimuth documentation (and the general understanding of CRISPR on-target scores, it looks like scores are expected to be between 0.0 and 1.0. Is this possibly a bug in the scoring system?
If it's helpful, this does not happen if the peptide percentage is set to 95 or less.
Thanks!
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FMicrosoftResearch%2FAzimuth%2Fissues%2F21&data=02%7C01%7Cjennl%40microsoft.com%7C1f8bbce5c0634723dc5a08d4df79c354%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636379161450199448&sdata=7GPe3j%2FrVz8ctSDCJQifbh%2FHv05T4sf5zklFPlSn%2FFo%3D&reserved=0, or mute the threadhttps://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAFj_NTH5QSWz5O61TE0tOvN5q4STggSlks5sWjjngaJpZM4Oyve_&data=02%7C01%7Cjennl%40microsoft.com%7C1f8bbce5c0634723dc5a08d4df79c354%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636379161450199448&sdata=Wv32WLQnFrW%2B7OAtSWRJYxcZA1yJZ6BKIqF7abyx2t4%3D&reserved=0.
jjc2718
Hello! We've noticed that occasionally, guides that we score will get a negative predicted score from azimuth. Here's an example:
Empirically, it looks like negative scores are more likely to happen with peptide percentages closer to 100. However, from the Azimuth documentation (and the general understanding of CRISPR on-target scores, it looks like scores are expected to be between 0.0 and 1.0. Is this possibly a bug in the scoring system?
If it's helpful, this does not happen if the peptide percentage is set to 95 or less.
Thanks!