Fig. 5: Generalizability of models.

We demonstrate the generalizability of models across varying dynamics parameters and dynamics equations. a–c Generalizability of models across varying dynamics parameters of mutualistic (a), gene regulatory (b), and neuronal (c) dynamics. For each dynamics, we randomly choose six of nine parameter settings (Supplementary Table 4) for generating training data. We employ each parameter setting to simulate node activity trajectories for 1000 perturbed networks via node removal from empirical networks and combine all data of different parameter settings, formulating 6000 samples. To facilitate the training process, we randomly select 1000 samples to construct the training set. Similarly, we employ each of the other three unseen parameter settings to simulate node activity trajectories for 200 perturbed networks via node removal from empirical networks and combine all data of different parameter settings, formulating 600 samples and randomly selecting 200 samples to construct the validation/test set. d–f Generalizability of models across different dynamics equations. We employ three parameter settings of SIS and inhibitory dynamics (Supplementary Table 5) to simulate node activity trajectories for random networks with (N, p) uniformly sampled from [30,60) and [0.05,0.25), respectively. We synthesize 1600 samples under each parameter setting of each dynamics and combine all generated samples, then randomly select 1600 samples as the training set. To construct the validation/test set, we use mutualistic (d), gene regulatory (e), and neuronal (f) dynamics, respectively. We also employ three parameter settings of each dynamics (Index 2–4 of Supplementary Table 4) to simulate node activity trajectories for 160 random networks with the same (N, p) distributions as the training set. For mutualistic, gene regulatory, neuronal, SIS, and inhibitory dynamics, we consider 10, 5, 11, 11, and 11 different initial conditions, respectively. Box plots depict the median (central line) of F1-scores (n = 35 with different random seeds), the first and third quartiles (box), whiskers extending to 1.5 times the interquartile range from the first and third quartiles, respectively, and outliers are represented as individual points.