Fig. 1: A schematic representation of dynamicGP.
a, For a single genotype with measurements for p traits at T timepoints, we seek to find a time-invariant best-fit linear operator A, which transforms the phenome, given by measurements for p traits, at time t into the phenome at time t + 1, according to \({{{x}}}_{{{t}}{{+}}{{1}}}={{{Ax}}}_{{{t}}}\). When organized into matrix form for all timepoints, we obtain a p × T matrix X. Two submatrices, X1 and X2, are offset by a single timepoint and are derived from X by omitting the measurements at the last and the first timepoint, respectively, and are then used to calculate A using equation (3) of the classical DMD method in algorithm 1 (Methods). b, When the time-resolved data are available for k genotypes, we obtain a p × p × k tensor whose elements can be treated as traits in heritability analyses and GP. c, The models for the operator entries are trained and then used to obtain predictions of entries, \({\hat{a}}_{{ij}}\) (red), for unseen lines, which are then gathered into a matrix that is then used to predict future timepoints for the line.
