Fig. 4: Feature encoding from artificial stimuli across ON bipolar cells.

a Schematic representation of the data augmentation process for classifying additional ROIs from calcium images of ON bipolar cell terminals, acquired from Grm6-Cre Ai148 mice. This classification leverages morphologically identified bipolar cells in conjunction with the scanning depth of the IPL. A semi-manual morphological segmentation approach is utilized, relying on the standard deviation of each pixel over time. To ensure diverse sampling, optimal k-means clustering is employed to group functionally similar segments. Classification hinges on the closest functional similarity and IPL profile match from the identified bipolar cell types. Following this, depth and functional probability maps are arranged based on the classification of the seven distinct bipolar cell types. Outliers are highlighted in black and excluded (<2% of ROIs, see Methods). b The encoding space, informed by both functional and morphological data, uses the first two coordinates from nonmetric multidimensional scaling (MDS1 and MDS2) for representation. The x and y axes are set to these coordinates. Individual bipolar cell types are differentiated by color-coding: BC5o (n = 45), BC5i (n = 40), BC5t (n = 45), XBC (n = 10), BC6 (n = 61), BC7 (n = 54), BC8/9 (n = 35). c Presents the distance matrix (the distance in the first three primary coordinates of the encoding space for artificial stimuli) for each bipolar cell type pairing. Asterisks in the upper triangle indicate statistical significance (* p < 0.05, ** p < 0.01, *** p < 0.001). Two-sided permutation tests were performed, p-values have been adjusted for multiple testing by controlling the false discovery rate (FDR) at a threshold of <0.05. d Left: The encoding space from (b) is overlaid with each ROI’s surround strength. Gradient colors encode the value of the surround strength. Top right: A probability distribution captures the angular relationships between subsampled paired vectors (90 pairs or 0.21%) and the collective paired vector angles for surround strength (as detailed in Methods). Bottom right: Each MDS coordinate’s variance contribution to surround strength is charted. e This panel parallels (d) but for response transience. Source data for this figure are provided as a Source Data file.