Opened 7 years ago
Last modified 7 years ago
#374 assigned defect
global 2DSA fitting method
Reported by: | demeler | Owned by: | gegorbet |
---|---|---|---|
Priority: | normal | Milestone: | future |
Component: | ultrascan3 | Version: | |
Keywords: | Cc: |
Description
The global 2DSA fitting method currently does not work correctly. It should be possible to fit multiple triples to a common 2DSA model. Such a fit should produce n+1 models, where n is the number of triples included in the fit. One possible algorithm would merge the results of individual fits into a common grid and then through iterative fitting find the smallest subset of solutes necessary to fit all data sets. Another possible algorithm would increase the size of the A matrix by concatenating all simulations for one grid and fitting them by NNLS to the b vector of concatenated experimental data sets. The final solution would represent the "+1" solution and could then be optimized to each individual data set by fitting it once more to each individual data set to get optimized solute concentrations for each solute, generating the n models for each data set.
Change History (2)
comment:1 in reply to: ↑ description Changed 7 years ago by demeler
comment:2 Changed 7 years ago by gegorbet
- Status changed from new to assigned
Replying to demeler:
After discussing with Emre, we will proceed as follows: Each global fit will be performed on data that have been pre-processed already with 2DSA to remove ti/ri noise, and fit the meniscus, iterative fit, etc. For each dataset, it is now important to get a scaling factor that will assure that each dataset has the same concentration, so a single model can be fitted to it. If this is not available from a previous fit, this scaling factor can be obtained by running independent 2DSA on each dataset.
Next, a big matrix A is constructed in which all simulated datasets are concatenated as long columns of A and b. the b vector elements are scaled to the same total concentration according to each individual dataset's total concentration, then the global model is created with a regular 2DSA/NNLS (perhaps with iterative fitting or with monte carlo).
In the next step, the global model (the n+1 model) is re-scaled to each individual dataset. There are 2 ways of doing this: 1. perform NNLS on the final global model against each individual dataset to get a best fit concentration for each solute. 2. rescale the concentrations for each solute for each individual dataset so that the total concentration matches the scaling factor determined in the first step.
this will generate 2n+1 models for the global fit. The first set of n models will have the relative ratios of individual solutes potentially changed and adjusted between individual datasets, the second set will ave these ratios constant. All 2n+1 models could be reported.