Comprehensive Analysis of the 16p11.2 Deletion and Null Cntnap2 Mouse Models of Autism Spectrum Disorder
Fig 3
NeuroCube found similar degrees of separation between 16p11.2 df/+ and Cntnap2 -/- as compared to their corresponding WT control littermates.
A & C: To build a 2D representation of the multidimensional space in which the two groups are best separated, we first find statistically independent combinations of the original features, pick the two new composite features (drf 1 and 2) that best discriminate between the two groups, and used them as x- and y-axes (see S4 Methods). As in Principal Component Analyses, these two axes represent the uncorrelated feature transformations that account for most of the variance. Each dot represents either a WT (blue) or a mutant (red) mouse. The center, small and large ellipses are the mean, standard error and standard deviation of the composite features for each group. The overlap between the groups is used to calculate the discrimination index, which indicates how reliably a classifier can be trained to discriminate between the two groups (the more overlap, the worse the discrimination). B & D: To estimate how likely it is that such separation is simply due to chance, the obtained classifier is challenged many times with correctly labeled samples (“WT” and “mutant”; see the green distribution) or with randomized labels (blue distribution). The overlap between these two distributions (in red) represents the probability of obtaining the observed discrimination by chance. A: At P30 the 16p11.2 df/+ standard deviation ellipse overlapped to a small extent with the WT control ellipse; B: An 89% discrimination between 16p11.2 df/+ and WT mice could be found by chance with p< .0003; C: At P30 the Cntnap2 -/- model separated well from the WT group; D: An 84% discrimination index can be found by chance with p< .003.