Develop Threshold Methods for Learn/Apply using Decoy Data

[jjacobson95]

Using profmark decoys, we can create thresholds based on what cosine similarity scores might be calculated between each family and randomized (decoy) sequences. These thresholds could be used to filter out predictions and increase the accuracy of Snekmer Learn Apply, particularly in distant or unrelated families.

Three proposed methods that utilize a threshold are as follows:

• Top Hit above Threshold
  • The top hit above the threshold is selected.
• Greatest Distance from Threshold
  • The value that is the greatest distance above the threshold is selected
• Balanced Distance with Threshold
  • The value selected is based on a weighting of the above two methods. Ideally this will optimized on a prediction basis.

A confidence metric will have to be tailored to each of the above methods as the current version does not take into account thresholds.