Fig. 4: AMNs growth rate predictions for E. coli gene KOs mutants.
From: A neural-mechanistic hybrid approach improving the predictive power of genome-scale metabolic models
An AMN model was trained on a set of 17,400 measured growth rates of E. coli grown in 120 unique media compositions and 145 different single metabolic gene KOs. a AMN architecture integrating metabolic gene KOs. This architecture is similar to Fig. 1c, except for a secondary input (RKO) for the neural layer, alongside the medium composition Cmed. The RKO input is a binary vector describing which reactions are KO. The custom loss function ensures that reference fluxes (i.e., the E. coli mutants measured growth rates) and mechanistic constraints are respected and that reactions experimentally KO have in Vout a null flux value. The neural layer comprised one hidden layer of size 500 and the model was trained for 200 epochs with dropout = 0.25, batch size = 5, and the Adam optimizer with a 10−3 learning rate. b AMN regression performance on aggregated growth rate predictions from a 10-fold cross-validation. The mechanistic layer used for this architecture was the QP solver. c Regression performance of classical FBA with scaled upper bounds for compounds present in the medium and setting the upper bound and lower bound to zero for reactions that are KO (having a value of 0 in RKO). d ROC curve of AMN results. We thresholded the measured growth rates (continuous values) in order to transform them into binary growth vs. no growth measures. e ROC curve of classical FBA results. The same thresholding as for (d) was applied. Source data are provided as a Source Data file (cf. “Data availability”).
