Figure 3: Performance comparison between mNODE and existing methods on real microbial community datasets.

Three metrics are adopted for comparing model performances: the mean SCC $\bar{\rho }$, the top-50 mean SCC ${\bar{\rho }}_{50}$, and the number of metabolites with SCCs larger than 0.5 $N_{\rho >0.5}$. All datasets^21,33–38 are randomly divided into training and test sets with the 80/20 ratio except for the PRISM and NLIBD dataset³³. a1-a3 Performance of methods after training on the PRISM and test on the NLIBD³³. b1-b3 Performance of methods on the data from lung samples of patients with cystic fibrosis²¹. c1-c3 Performance of methods on the data from soil biocrust samples after 5 wetting events³⁴. d1-d3 Performance of methods on the data from fecal samples of children at age3^35–38. e1-e3 Performance of methods on the data from blood plasma samples of children at age 3^35–38.