Fig. 3: Benchmarking growth rate predictions by AMNs with experimental measurements.
From: A neural-mechanistic hybrid approach improving the predictive power of genome-scale metabolic models
In all panels, the experimental measurements were carried out on E. coli grown in M9 with different combinations of carbon sources (strain DH5-alpha, model iML1515). Training and 10-fold stratified cross-validation were performed three times with different initial random seeds. All points plotted were compiled from predicted values obtained for each cross-validation set. In all cases, means are plotted for both axes (measured and predicted), and error bars are standard deviations. For the measured data, means and standard deviations were computed based on three replicates, whereas for predictions, means and standard deviations were computed based on the 3 repeats of the 10-fold cross-validation. a Architecture and performance of AMN-QP. The neural layer (gray box) is composed of an input layer of size 38 (Cmed), a hidden layer of size 500, and an output layer of size 550 corresponding to all fluxes (V) of the iML1515 reduced model. The mechanistic layer (green box) follows the neural layer and minimizes the loss between measured and predicted growth rate, as well as the losses of the metabolic network constraints. The model was trained for 1000 epochs with dropout = 0.25, batch size = 5, and the Adam optimizer with a 10−3 learning rate. b Architecture and performance of AMN-LP. This model hyperparameters are identical to those of (a). The neural layer computes the initial values for the 550 reaction fluxes (vector V), the initial values for the 1083 metabolite shadow prices (vector U) are set to zero. c Architecture and performance of the AMN-Wt architecture. The model hyperparameters are those of the previous panels and the size of the Wr matrix is 550 × 1083 (sizes of V and U vectors). Source data are provided as a Source Data file (cf. “Data availability”).
