Figure EV5. CODEX identifies different sex‐ and light‐dependent Drosophila movements.

Male and female Drosophila movements in a tube were recorded over 12 h, in presence or absence of light, and reported as univariate time series (Fulcher & Jones, 2017). A CNN was trained to recognize the combination of sex and light conditions based on the individual movements (see Materials and Methods and Appendix Note 5).

1. t‐SNE projection of the CNN features for the training, validation, and test sets pooled together. Diamonds indicate the 10 top prototypes for each class.
2. The top prototype trajectory for each class.
3. Comparison of autocorrelation levels and spectral flatness between the 20 top female prototypes and 20 top male prototypes against non‐prototype trajectories. Female (resp. male) prototypes were obtained by pooling the 10 top prototypes of female (resp. male) drosophila under both day and night light conditions. Boxes indicate the upper and lower quartiles, the central band indicates the median and whiskers extend to individuals up to 1.5 interquartile away from the median. Medians of the distributions were compared with two‐sided Wilcoxon's tests; ***P‐value < 0.01. n = 574 for non‐prototype female trajectories; n = 524 for non‐prototype male trajectories.
4. Autocorrelation level and spectral flatness against the probability of trajectories to be classified as a female class. The latter probability is the sum of the probability of trajectories to belong to the female‐day and female‐night classes. The resulting sum was ranked, such that rank 1 corresponds to the probability of the trajectory that has the lowest prediction outcome to be classified as female. Autocorrelation levels were computed by averaging autocorrelation values for lags ranging from 1 to 10 time points.