Modeling:Idl/Base

From QuB

Jump to: navigation, search
Prev: Modeling:Amps Outline Next: Modeling:MIL


Idl/Base (IB) does four related things:

  • customizable idealization with the best elements of SKM and Amps
  • constraints on amplitudes and noise see this forum post
  • idealization while tracking baseline drift, using a Kalman filter
  • idealization of Staircase data
  • onscreen visualization (movies) of idealization, to help understand the algorithm

The details are explained in Dr. Lorin Milescu's thesis.

Contents

Idealization

Idealization consists of one or more iterations of

  1. find the most likely state sequence (MLSS) given the model and sampled data
  2. use the MLSS to re-estimate model parameters

The MLSS can be computed by

Viterbi
google:Viterbi
Gamma
the most likely state s at time t maximizes Γs,t as computed by Baum-Welch
Forward-Viterbi
google:Forward-Viterbi

Re-estimation is optional. Repeated re-estimation and iteration can improve the idealization, but if the model parameters are already optimal, it's much faster to just perform step 1 once. Your choices are

None
skip step 2 and don't iterate
Seg K-Means
as in SKM
Baum-Welch

Baseline

Idl/Base can track a drifting baseline while idealizing, using a Kalman filter. It can also correct the baseline by adding baseline nodes. Check "bline apply" at the top to subtract the node-baseline from the data.

The theory is an extension of the Viterbi and Forward algorithms. When evaluating P(s1 at t, given s0 at t-1), we calculate the emission probability using (datat - predicted next baseline point from s0,t-1) and bring the filter forward with (datat - amps).

Staircase

A staircase process has jumps in only one direction. An example is fluorescence position data from molecular motors such as actin.

TODO: how is the model interpreted?

Movies

Idl/Base can show the progress of idealization and baseline tracking as a movie in the Data window. If you're trying the movie option, use a short selection of data.

Properties

Data channel index which A/D channel contains the data to idealize, typically 0
Replace Idealization completely remove any idealization, even in parts of the file that are not part of the data source. Otherwise, idealization outside the Data Source is kept.
Quiet output prints significantly less information to the Report window
Every iteration prints information about each iteration to the Report window
Data source
Pre-process data none: use raw data
as displayed: use the same baseline correction and filtering as the Data display
as such: apply the specified filter and baseline
Make movie show a movie of the idealization's progress in the Data window
Save bitmaps saves the movie as a sequence of .bmp files somewhere
Show titles
Max frame count
IDL
BX
BP
Colors
Re-estimation described above
A -> Q after re-estimating the A matrix (sampled transition probabilities), calculate rate constants (Q matrix) which give rise to that A matrix, and update the model
Every iteration update the model's rates from the A matrix every iteration, or just at the end
Fix kinetics
Fix amps
Fix noise
These checkboxes do nothing since we switched to the amplitude constraint system
Max iter repeat (find MLSS, re-estimate params) at most this many times
LL conv stop iterating when the log-likelihood improves by less than this much
Most likely sequence described above
Baseline Track track baseline drift with a Kalman filter
Baseline Std baseline drift from sample to sample is modeled by a Gaussian distribution
Baseline Correct add baseline nodes to the Data corresponding to the tracked baseline
Min dy add a node when the baseline changes by more than this many A/D units
Back Kalman
F-K Count
F-B-K Count
Reest params
Staircase
Max jump
Right only
Global est
LL add Jump
B-W reest
BaseJumpV
UpdateRefState
every
Step Amp reest.
Steps
Init
SSj / SSm
Weighed
Cons. re-est +/-
Merge dwells if
dA < [frac]
Split dwells if
dA > [frac]
Kinetics re-est
Init
kf
kb
Step Std reest
Initial value
Min Gamma
Weighed
Amp(0) initial value
Meas std reest
Meas std Initial value

Results

In the textual Report window: exhaustive details of the computation

In the Results window:

Summary:

amp the (re-estimated) mean amplitude of each class
sd the (re-estimated) standard deviation of each class

Segments (and Select, Criteria): all the measurements of each segment/selection

Models: the model used for idealization

Histograms: An amplitude histogram for each segment/selection, overlaid with the Gaussian distribution (amp, std) of each class. You can quickly identify segments that were mis-idealized by looking for histograms whose distribution curves don't match the histogram bars.


See Also


Prev: Modeling:Amps Outline Next: Modeling:MIL
Personal tools