In search of early cortical mechanisms for color: individual variability in steady-state VEP amplitudes for hues sweeping around the isoluminant LM and S cone-opponent plane

From retina to V1, chromatic mechanisms fall along two cone-opponent ‘cardinal’ axes (L-M, S vs. L+M). Is this coding propagated to early cortex? To explore possible hue-selective cortical responses, we examined individual variability in steady-state visual evoked potentials of 16 participants, using a flickering checkerboard whose color swept around the isoluminant hue circle at three chromatic contrasts (Kaneko et al., 2020). For each hue, intra-individual correlations (r) were strongest with hues at nearby angles, consistent with multiple channels. But there were multiple excitatory and inhibitory sidelobes (+, - correlations) at distant angles. Remarkably, at the highest chromatic contrast, six sidelobes were at 0 (red, +[L-M]), 30 (magenta), 120 (blue), 180 (teal, -[L-M]), 240 (green), and 300 (yellow) degrees. Nonmetric Multidimensional Scaling* of the dissimilarity matrix (1-r) estimated four significant components. The first 2 components’ loadings showed multiple + or – peaks that aligned closely with the 6 sidelobes: red vs. blue and green; teal and magenta vs. yellow and blue. The 2 additional components added 4 colors: 90 (violet, [+S]) vs. 150 (blue-green) and yellow; 270 (lime, [-S]) and magenta vs. blue-green, blue and 330 (orange). (Some components differed/shifted for ½, ¼ contrasts). If these 4 exploratory, unvalidated components represent cortical hue-selective mechanisms, then they show neither classic cone- nor unique-hue opponency, nor complementarity, nor simple narrowly tuned color channels. Rather, the 4 ‘mechanisms’ subserve ten physiologically primary hues, and each primary hue has +, -, and Ø interactions with the other 9 hues. Possible implications of a new chromatic coding model are discussed for chromatic induction and assessment of singe-cell tuning. (*NMMDS, compared to PCA/FA, is nonparametric [i.e., no linearity, metricity, or multivariate normality required of data or underlying components]; doesn’t force ‘simple structure’ on data generated by broadly tuned overlapping mechanisms; and needs fewer components to explain variability).