Figure 5

Discriminators trained on MaxCal behavioral descriptors perform better than shape alone. (A–C) Success rates for k-nearest neighbors based on the shapes of cells alone. Some number of samples are taken from the same condition and individually classified using the first 1 (A), 2 (B) or 3 (C) shape PCs, with the most commonly chosen classification used to identify the group. Classification is performed on both untreated vs BPB/LY treated (blue) and parental vs mhcA⁻ (green) data. (D) A similar k-nearest neighbor classification to (A–C), but with distances between neighbors based on the values of the two Lagrange multipliers that best separate the training data. (E) Degree of correlation between PC 1 and PC 2 across a number of simulated cell trajectories. Three lines are shown, corresponding to the data (black), a non-Markovian (red) master equation, in which transition probabilities depend on the last transition made, and a Markovian (blue) master equation, in which they simply depend on the current state. The Markovian model (without memory) lacks the skew toward negative correlations seen in the data. The non-Markovian model (with memory) recovers these.