Extended Data Fig. 5: MDSINE2 accurately recovers dynamical ecosystems from semi-synthetic time-series data (growth rates and perturbations).
From: Learning ecosystem-scale dynamics from microbiome data with MDSINE2
(a) Spearman correlation for growth rate strength prediction. (b) AUC-ROC for predicting taxa perturbation presence/absence. (c) Spearman correlation for predicting taxa perturbation strengths. MDSINE2 and MDSINE2−M outperform all other methods in all scenarios except for the ¼ density temporal sampling regime. For all scenarios n = 30 simulations, 10 different seeds for the initial conditions used to generate trajectories and 3 seeds for measurement simulation (read abundances and qPCR values). Boxes denote interquartile region with a line for the median. Whiskers denote 95% interval. Statistical significance tests were performed within each time-series down sampling scheme for the four methods either trained on all available data or reads/relative abundances. All comparisons are significant (p < 0.05) with BH correction unless denoted with “n.s.”. All p values provided in Supplementary Data Table.
