Two-photon calcium imaging

Two-photon imaging datasets were acquired during days 1 (baseline), 4 (after learning), and 7 (after relearning) for control experiments, where 200–300 neurons from five mice were analyzed. Motion artifacts were corrected for by a technique based on nonlinear optimization and discrete Fourier transform (DFT) in low noise imaging⁸⁴ for the case of uniform motion artifacts, where the quality of image registration was assessed using normalized root-mean-square (NRMS) between reconstructed and reference images⁸⁵ using the first image frame as the reference. Alternatively, we used a template matching method that split the field-of-view into spatially overlapping patches according to user-determined dimensions, registered corresponding patches of the template separately and then merged the registered subpatches to each other⁸⁶ for the case of nonuniform motion artifacts. Image registration quality was measured by the image crispness defined as the Frobenius norm of image gradient vector and image magnitude.

To extract the neuronal change in fluorescence, we used a method that automatically identified ROIs (including spatially overlapped ones), denoised signals, and when comparing the ITI period to the immediately preceding reward period (Fig. 3f; Supplementary Figs. ¹⁰, ¹¹) deconvolved signals⁸⁷ with open source and adapted MATLAB code (MathWorks, MA, USA). Briefly, this method uses constrained non-negative matrix factorization (cNMF) to isolate spatially and temporally independent fluorescent signals, approximating a parametric model for continuous timeseries calcium transients as the impulse response of an autoregressive process, and then estimates the spiking signal from the sparsest non-negative neural activity signal. This method can denoise the spatiotemporal imaging set and model the background activity in each image frame by averaging the spatiotemporal background over ROIs. The temporal trace of each ROI was expressed as ΔF/F, raw fluorescence trace divided by background activity. The default parameters were used with few exceptions (the order of AR process p was set to 2, temporal downsampling factor “tsub” was set to 4, spatial downsampling “ssub” was set to 2). Between 50 and 70 ROIs corresponding to the somata of neurons were identified per animal per session using this approach after manual confirmation.

To follow the changes in the encoding pattern of individual cells across days during learning and reversal learning to determine how neurons remapped response categories across non-encoding, single-, double-, or multidimensional encoding, ROIs were also automatically identified using suite2p⁸⁸ simultaneously across days 1, 4, and 7 for four out of the five mice used in the control group (the remaining mouse was excluded due to the field-of-view on baseline day not matching precisely to the following chronic imaging days).