Christian RohrsenView Profile
Mean trace of the positive control in the Joystick to get to see what are the overall dynamics and maybe to get an idea what might be the best score to pick.Here the standard deviation of the flies along the time axis. This is just to see if all the flies have more similar phenotypes with each other or not at each time.
This is to see if the flies have less degrees of freedom at any segment by measuring the standard deviation at each segment. There does not seem to be any effect. Although this might be mixed with the wiggle scores. I think measuring entropy is a better measure.
All the same plots as above but for TH-D’, the interesting line from the screen.
This are the performance indices for the different models performed to estimate the valence of the dopaminergic clusters. AIC: Akaike Information Criteria; BIC: Bayesian information Criteria; LogLikelihood: log Likelihood estimation
lm: linear model
+ int: taking double interaccions into consideration
b lm: bayesian linear model with bayesglm function
b lm MCMC: bayesian linear model with MCMCglm function
nlm: nonlinear model with lm function with splines fitted
b nlm: splines fitted to each cluster and MCMCglm function
GAM: general additive model with gam function
Adding double interactions seems to produce better models, nonlinearities also make models better and frequentist also. To me it seems like this data might be noise and therefore adding interactions, nonlinearities and frequentist methods is just fitting the noise better (overfitting) and that is why I get better scores with them. In addition, care needs to be taken since I use different functions that calculate the model performance scores differently (although the formulas are theoretically the same for all!)
This is the correlation from the T-maze experiments from Gaia and Naman. Neither ranked nor regular correlation show any significant effect. This means that these effects seem to be random, at least for most of them, is this an overfitting result?
I would say blue 1 is a line that was negative for all the tests I have so far seen. So this might be an interesting line. What to do next?
I would unblind the blue1, which is TH-D’. It was shown to be required for classical conditionning in shock and temperature learning (Galili et al. 2014). Another interesting observation is that th-g4+th-g80 seems to have like zero PI scores in all of the experiments (Naman and Gaia in the Tmaze, Joystick and Y-mazes). So could it be that all of these neurons have indeed a meaning, but is depending every time in the context?? Maybe Vanessa Ruta´s work might be interesting for that.
This is the linear model with its statistics
This is the linear model adding interactions. It is perfectly possible to have interactions between neurons, kind of what occurs with olfactory processing where ORNs activated alone or in different combinations have completely different meanings
I uploaded in slack the bayesian linear model with interactions. For any reason, it does not let me upload it now to the website
I am trying one of the ways of nonlinear models: GAM (Generalized additive models). Here one fit splines to the effects to certain degrees of freedom.
This is the kind of bar graphs I thought I could use for all the plots.
Dopaminergic clusters are differently targeted by the different Gal4s. Some of the express faintly, others stronger. Here I try to see if the dose-response curve (or expression-PI curve) seems to be linear or not. Here I put two examples from the 17 clusters, where the first two seem to have nonlinear curves, with and optimal expression level, and the last two seem to have a linear response curve.
This will be important for the modelling in order to decide to make a linear/nonlinear model. Down below I show the results from a linear model and it´s statistics. From Aso et al. 2012, one could see that activating the lines with TrpA1 shows a linear response curve. But in this case it does not necessarily seem to be the case. Therefore, light intensities might have an effect, as well as the expression level, and conclusion needs to be taken carefully.
In addition it is difficult to calculate this for all the clusters with just one single light intensity test, because not all clusters are expressed in several Gal4s to different level, so that we can estimate from there. So for the interesting lines we might need to make several experiments at different intensities, and see the dose response curve.
Confocal image MAX stack of one of the brains at 20x
and at 40x
In this link we have a video of a 3D stainning pattern zoomed_CC
Axel stainning from both drivers together. I would say it really contains both driver lines.
This are my stainnings at the fluorescence microscope (no confocal). This is to show that in all of the 10-12 brains I have looked at, they all had the c232 pattern present
In addition, they had many more neurons outside from the central complex which I believe belong to the c105-G4 line. This is my only proof to show that c105 is also present, since the R1 neurons seem to be hidden when R2 neurons are stained.
After performing EMD to 6 fly traces of 20000 data points (that is 1000 sec flight) for each group (tntXwtb; c105;c232>tnt; c105;c232Xwtb). This data size was chosen to reduce computing time of the SMAP procedure. The EMD decomposes the trace into different time scales in nonstationary data. It seems that the nonlinear behavior occurs at the first IMF (the fastest time scale) and a bit in the second IMF. The potential conclusion to this is that the behavior of the the fly is only unpredictable at the fast movements whereas slow movements are very predictable. Nevertheless, to be cautious it could be that this fastest timescale is just noise, and that this noise is nonlinear. I would say that there is no difference at any time scale between groups (pay attention to the different ranges in the Y-axes), so the ring neurons R1, R3, R4d do not have any effect.
As a groundtruth I have used the same analysis pipeline for the traces in the uniform arena from Maye et al. 2007. Here the effect is even more pronounced at the fastest time scales. So I will conclude that this is real fly behavior and not noise that is shared among both setups: the Ping pong ball machine and the torquemeter.In order to gain more insights into the underlying flight structure I took one random flight trace to explain a few observations. The x-axis is the theta (that actually goes from 0-4 in steps of 0.2 and therefore we see the 21 points), in the y-axis is the correlation of the prediction to groundtruth. We see that IMF has a bigger slope, but not only that, also that its prediction correlation is around 0.88, whereas lower timescales prediction is basically perfect. That is, fast time scales are not only more nonlinear but also less unpredictable. This pattern is repeated in every fly measured