search

stephlj / slopey

Quantification of single-molecule FRET trajectories containing fast, non-instantaneous transitions
1 stars 0 forks source link

Adjust prior? #2

Closed stephlj closed 8 years ago

stephlj commented 8 years ago

Here's some real data I thought would be pretty obvious re: where the slopey bits are: ("... Spot1_320") model_fretplot_snf2h103nmatp1mmhighfps_2015_dec04_spot1_320

Doesn't seem to work very well though, here's the output of the slopey analysis:

screen shot 2016-01-07 at 12 15 52 pm
mattjj commented 8 years ago

Did the 1 channel version work better?

mattjj commented 8 years ago

I'm away from wifi but a direction to try would be 

slopey_time_hypers = 1., 3.  # exp prior with mean of 1/3 sec
flat_time_hypers = 1., 1./10  # exp prior with mean of 10 sec
stephlj commented 8 years ago

Looking back at closed issue #1, I think the one-channel version did work better. I'm going to checkout an older version and confirm that.

mattjj commented 8 years ago

That would be a bit more complicated. It could mean that green is a non-affine function of red. Or that there's a bug.

We should tag a one-channel version. Or find a way to make them coexist. That wouldn't be too hard.

mattjj commented 8 years ago

Wait, is #1 the same data?

The 2ch thing didn't affect multiple slopey bits. The thing in question here is if the 1ch version works better on this same data sequence.

stephlj commented 8 years ago

No, #1 isn't the same data as here. Here's the data that works well in #1 with the 1ch version, but doesn't work with the 2ch version:

screen shot 2016-01-07 at 1 58 12 pm
stephlj commented 8 years ago

Neither 1ch nor 2ch works well with what I thought was an "easy" trace (the images in the first comment in this issue). Changing the prior changes the results a lot for both 1ch and 2ch but I couldn't find a prior that made the "easy" data work with either version.

stephlj commented 8 years ago

wait hang on, that's not the same data either, even though it's supposed to be ...

stephlj commented 8 years ago

Ok. The 1ch version that worked, worked on trace 115, starting from frame 250. If you run the 2ch version on that trace, cropping the first 250 frames, it does work:

screen shot 2016-01-07 at 2 08 54 pm

That works well at ID'ing what I think are slopey bits, and the duration plot looks reasonable, with the peak around 0.3 where I was expecting it. However, now that it's working, I'm realizing a problem with the transformed green data. This particular trace photobleaches before the reaction completes, meaning the red doesn't get all the way to zero, and the green shouldn't get all the way to where the red starts. Does your model insist that red gets to zero by the end and green has to get to the red's starting level ... ?