Figure 2: Model comparison and validation of mNODE on synthetic data generated by the microbial consumer-resource model.

The predictive performance of mNODE is compared to other methods through three metrics: a1 the mean SCC (Spearman Correlation Coefficient) $\bar{\rho },$ a2 the top-5 mean SCC ${\bar{\rho }}_5$, and a3 the number of metabolites with an SCC larger than 0.5 $N_{\rho >0.5}$. b1-b3 The predictive performance of mNODE as a function of training sample size when the nutrient sampling probability (i.e. the fraction of nutrients externally supplied) $p_{\mathrm{n}}=0.6$ and the species sampling probability (i.e. the fraction of species introduced) $p_{\mathrm{s}}=0.5$. c1-c3 The predictive performance of mNODE when the training sample size is 240 and the species sampling probability $p_{\mathrm{s}}=0.5$. Solid lines with circles are predicted results when nutrient supply rates (i.e. diets in the microbial consumer-resource model) are not included in the input of mNODE. Dashed lines with triangles are predicted results when nutrient supply rates are included in the input of mNODE.