These are chat archives for openworm/ChannelWorm

27th
Sep 2015
Richard C Gerkin
@rgerkin
Sep 27 2015 16:06
OK, here
Stephen Larson
@slarson
Sep 27 2015 16:06
@rgerkin have you logged into http://channelwormdjango-channelworm.rhcloud.com/ recently?
@rgerkin there should be digitized data from there
Richard C Gerkin
@rgerkin
Sep 27 2015 16:06
@slarson Is there anything there above and beyond what is accessible through the Django ORM locally when I am using the ChannelWorm repo?
Stephen Larson
@slarson
Sep 27 2015 16:07
@sumwor have you made any progress or still finding your way?
@rgerkin I believe there is data up on that cloud instance that has not been synced back to the ChannelWorm repo. @VahidGh -- is that right?
Richard C Gerkin
@rgerkin
Sep 27 2015 16:09
Unless the rhcloud site has a REST API, I think it will be easier to use the data available in the ChannelWorm repo, so I can create IPython notebooks with the whole testing workflow self-contained.
@slarson @VahidGh My first goal would be to pick 3-5 figure panels that we want the model to replicate and think the model might be capable of replicating, and then I can work on figuring those out.
summerworm
@sumwor
Sep 27 2015 16:10
@slarson yes a little progress, actually I transfer to UCI this quarter so I am still trying to get used to the school and lab works. It is way different than things back in China! I'll keep working ASAP
Stephen Larson
@slarson
Sep 27 2015 16:10
@sumwor UCI as in Irvine?
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:11
@slarson, I synced the last updates recently
summerworm
@sumwor
Sep 27 2015 16:11
@slarson yeah
Stephen Larson
@slarson
Sep 27 2015 16:11
@VahidGh OK great -- can you link @rgerkin and I to where the new data reside when they are in GitHub? I don't think I'm familiar with that part of the project
@sumwor Oh very cool. I'm in San Diego. We should try to meet up some time!
summerworm
@sumwor
Sep 27 2015 16:12
@slarson that will be wanderful!
Richard C Gerkin
@rgerkin
Sep 27 2015 16:13
I count 22 figure panels in http://channelwormdjango-channelworm.rhcloud.com/ion_channel/graph/; I'd like to focus in on 3-5 of those (having only done one of them so far), but I don't know what is thought to be important and reliable.
Stephen Larson
@slarson
Sep 27 2015 16:15
@rgerkin @VahidGh I'd vote for working on the SLO-2 ones, but @VahidGh perhaps you can say which one makes most sense given what you've been most successful doing the optimization for?
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:15
@slarson, This could be the same as what we done once here
Richard C Gerkin
@rgerkin
Sep 27 2015 16:16
@slarson @VahidGh yes, that's what I want to replicate for more experiments.
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:16
SLO-2 is in priority but is from a more complicated family of K channels
Stephen Larson
@slarson
Sep 27 2015 16:17
@VahidGh Gotcha-- what's simpler?
EGL-19 like we started?
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:18
SLO-2 would be OK, btw
Richard C Gerkin
@rgerkin
Sep 27 2015 16:18
I don't mind a complicated channel as long as jNeuroML doesn't choke on it
Stephen Larson
@slarson
Sep 27 2015 16:18
hehe
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:18
We can start with a simple model and then adding the other parameters
Stephen Larson
@slarson
Sep 27 2015 16:19
@VahidGh is it possible yet to populate the ion channel models screen? https://channelwormdjango-channelworm.rhcloud.com/ion_channel/ion_channel_model
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:20
It's possible using command lines, but still it's not been integrated with the UI
The server is not powerful enough to handle optimization tasks
Stephen Larson
@slarson
Sep 27 2015 16:21
Gotcha. Have you been storing optimized models somewhere in the repo via command lines yet? I'm again, not as familiar about where they live
Richard C Gerkin
@rgerkin
Sep 27 2015 16:21
This looks like a place where integration with NeuroML-DB could be useful
We want to store a bunch of ion channel models, provenance, etc. there, too
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:23
@slarson, there are some examples here I was working on
@rgerkin, yeah
Stephen Larson
@slarson
Sep 27 2015 16:24
@VahidGh Ok Cooooool!!
@VahidGh So you haven't had to go through NeuroML channels yet to do that optimization, yeah?
Because yeah, @rgerkin -- that could be interesting to add a step to convert to NeuroML channels & combine with NeuroML-DB. I'm not as familiar as to where that project is right now
Richard C Gerkin
@rgerkin
Sep 27 2015 16:26
What is channel format now?
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:26
Actually I'm doing the optimization and then passing the parameters to NeuroML file generators through this
Stephen Larson
@slarson
Sep 27 2015 16:27
OIC so that's a subsequent step within the fitter framework!
Thanks for filling me in & sorry for my ignorance!
I think I got confused because you aren't writing them out into the repo
So that would be an option to add, potentially, right?
even though, hmm, looks like you are intending to do something like that here: https://github.com/openworm/ChannelWorm/blob/master/channelworm/fitter/examples/SLO-2.py#L277
Richard C Gerkin
@rgerkin
Sep 27 2015 16:29
@VahidGh Looking at that code, are you just extracting channel parameters reported in the papers and then building NeuroML files with those parameters?
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:29
@slarson, exactly :D
@rgerkin, Nope, I'm extracting the parameters from those figures
Richard C Gerkin
@rgerkin
Sep 27 2015 16:31
Oh, but are they directly determinable from the figures or is there a modeling step there?
Things would be much easier if all data was reported in tables (in addition to figures).
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:32
I'm getting the digitized data, then starting the optimization and testing different parameters, then storing the best model within a nml file
Richard C Gerkin
@rgerkin
Sep 27 2015 16:33
"Best" in reproducing the figures from which you extracted the parameters, or some other figures?
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:33
Yeah, there are some TODOs that was supposed to be completed by some "helping hands" :D
Richard C Gerkin
@rgerkin
Sep 27 2015 16:36
@slarson @VahidGh Alright, I'm going to dive into an SLO-2 channel figure, maybe by analogy to the EGL-19 one I already did, and try to write a testing notebook for it.
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:36
I'm really busy with the other milestone, and had some great progress, but lacking some super computer for the tasks such as grid searches I'm doing for feature selection made the work a bit slow. Universities are getting open by this week in Iran, and I hope I could improve the speed by having some powerful computers in our lab
Stephen Larson
@slarson
Sep 27 2015 16:37
@rgerkin ok awesome!
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:38
I'm trying to find some relationship between biophysical and biochemical features (like these of ion channel proteins and their kinetics
So far I could reach an estimation of the kinetics with the MAE of about ~9 mV
And decreasing the error day by day by playing with those features
The Scikit learn is a cool package, and getting master on it :D
Richard C Gerkin
@rgerkin
Sep 27 2015 16:41
@VahidGh Scikit-learn is pretty great!
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:44
@rgerkin, @slarson, Oh I forgot to mention this great work, recently launched: http://icg.neurotheory.ox.ac.uk/
And thanks to Dr. Crook for the clue
I talked to the guys behind this
To see if we can include my idea
They have some similar idea based on clustering techniques
And their publication is under review
My plan is to find those parameters from protein sequences and then for each channel, find the most similar model, and then include my parameters into the channel model
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:50
unfortunately they don't have models for C. elegans, but we can use my approach to find out the best model for the channels we don't have patch clamp data for
Stephen Larson
@slarson
Sep 27 2015 16:50
@VahidGh Wow just been looking on that site
that seems extremely relevant!
Richard C Gerkin
@rgerkin
Sep 27 2015 16:51
@VahidGh Oh wow, surprised I haven't seen that
Stephen Larson
@slarson
Sep 27 2015 16:51
Me too!
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:52
The first version launched just a month ago
@rgerkin, @pgleeson, is there any script to convert a NEURON model to the NeuroML file?
Richard C Gerkin
@rgerkin
Sep 27 2015 16:54
@VahidGh I think NEURON (the program) has NeuroML export. I don't know how well it works. Padraig would have the best word on this
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:55
That would be great
An interesting part is that they've provided a REST API for accessing raw data
Richard C Gerkin
@rgerkin
Sep 27 2015 16:56
@VahidGh You mean the ion channel site you just linked?
Vahid Ghayoomie
@VahidGh
Sep 27 2015 16:56
yep
This way we can estimate each channel model by the predictor I'm working on, then using those APIs, find the most similar model for the ion channel of interest, and then convert it to NeuroML
Vahid Ghayoomie
@VahidGh
Sep 27 2015 17:01
A problem with ion channel kinetics (especially gating) is that there could be two channels from the same family and subfamily but with different kinetics in different species
And vice versa for channels with different families but the same kinetics
So, we can not just rely on families
By combining two approaches, we can estimate gating kinetics from the protein and structural properties, and then find the other parameters (such as activation/inactivation and their powers) from similarity/family based approaches, and then having the best estimated model
Stephen Larson
@slarson
Sep 27 2015 17:22
Cool idea
Richard C Gerkin
@rgerkin
Sep 27 2015 17:22
Yes, I like it
@VahidGh How would you search by model similarity? Does the API let you do that?
Vahid Ghayoomie
@VahidGh
Sep 27 2015 17:27
@rgerkin, using the browser we can limit the models we are interested in, and then using the APIs getting the raw data for each plot, and then by comparing with our estimations, we can find the best match
I also am waiting for their publication to see what are the features for the kinetics similarity , etc
We can also use the same approach for the second part of our approach
Richard C Gerkin
@rgerkin
Sep 27 2015 17:29
Oh so you are selecting models based on similarity to your models' output plots, or to your digitized data plots, or to model parameter similarity?
@VahidGh
Vahid Ghayoomie
@VahidGh
Sep 27 2015 17:30
This is what they intend to do, and the aim of the project
Richard C Gerkin
@rgerkin
Sep 27 2015 17:30
@VahidGh I mean what methods are you taking:
1) Select based on similarity to your model's simulation plots
or 2) Select based on similarity to data plots of interest
or 3) Select based on similarity of model parameters to your fitted model's parameters
Vahid Ghayoomie
@VahidGh
Sep 27 2015 17:34
Yeah, the first one -- I'm estimating the kinetics, then generating some putative model, and then comparing mine with the most similar one within the same family to complete my model based on what there have been generated for the same family
Richard C Gerkin
@rgerkin
Sep 27 2015 17:34
Cool
Vahid Ghayoomie
@VahidGh
Sep 27 2015 17:34
They don't work with the parameters, they just store traces, and the similarity is based on traces
They've also normalized currents, so there is no problem with max conductance
We just need to include the models within the cells, and then optimize parameters like conductance with some approach taken by @pgleeson in C302