Extended Data Fig. 5: Generation and analysis of GLMs for DA and Ach signals.
From: Dopamine and glutamate regulate striatal acetylcholine in decision-making
a, GLM workflow. Input variables are convolved with their kernels, with each time step consisting of a separate β coefficient fit by minimizing a cost function. The convolved signals are summed to generate the reconstructed signal. b, Evaluation of GLM performance. The original dataset is parsed into training and test sets. The GLM model is generated from the training set, and its performance is evaluated with MSE and R2. To generate confidence intervals for the MSEs (mock plot shown), the data are resplit ten times for f and three times for e and i. c, Kernels and reconstructed DA signals for the base GLM. The average photometry signals with bootstrapped 95% C.I. and the average kernels ± s.d. are depicted (n = 8 mice). d, Kernels and reconstructed Ach signals for the base GLM. Data are depicted as in c (n = 9 mice). e, Different hyperparameter sweeps over regression models – OLS, lasso regression (L1), ridge regression (L2) and elastic net (L1 + L2), and effect on indicated MSEs of the DA and Ach base GLMs (DA: n = 8 mice; Ach: n = 9 mice). Box plots are displayed as quartiles (25%, 50% and 75% percentiles) with 1.5 × interquartile range for whiskers and outliers marked as points outside this range. f, The effect of omission (−) or inclusion (+) of the indicated input variables on GLM performance, as measured by the effect on indicated MSEs (DA: n = 8 mice; Ach: n = 9 mice). The box plots are displayed as in e. g, Kernels and reconstructed DA signals for the history GLM. Data are depicted as in c (n = 8 mice). h, Kernels and reconstructed Ach signals for the history GLM. Data are depicted as in c (n = 9 mice). i, The effect of different hyperparameter sweeps for history GLMs. Data are depicted as in e (DA: n = 8 mice; Ach: n = 9 mice).
