Fig. 6
Computer models predicting the optimal pathway architectures to maximize the naringenin production titer in Escherichia coli. A Ordinary least squares (OLS) regression plot comparing Leave-one-out (LOO) model predictions of the training data to the actual measured titer. The final model holds an R2 of 0.93 and p = 1.48*10–7 (Additional file 1: Fig. S4). Also the top six predicted producers with pathway architectures and derived consensus architecture are given. B Partial least squares (PLS) regression plot comparing LOO model predictions of the training data to the actual measured naringenin production titer. The final model [2] latent variables] explains 78.92% (= R2) of the naringenin product titer variance by using 38.82% of the predictors’ variance. Also the top six predicted producers with pathway architectures and derived consensus architecture are given. C Machine-learning workflow developed by Zhou et al. [30] to optimize a biosynthetic pathway, here applied for naringenin biosynthesis. 1000 iterations of the ANN (artificial neural network) are trained with random initial weights. In each iteration the titers for the complete search space are predicted and the ten best producers for iteration i are stored. The frequency (f) of the occurrence of each unique pathway in the Top10 lists of all iterations is calculated and a 0.5*fmax threshold is set to select the most promising architectures. (P_X: promoter driving expression of enzyme X; CDS_X: enzyme variant of enzyme X; Rg: Rhodotorula glutinis; Fj: Flavobacterium johnsoniae; Pc: Petroselinum crispum; Ph: Petunia hybrida; Gh: Gerbera hybrida; Ms: Medicago sativa; TAL: Tyrosine ammonia-lyase; 4CL: 4-coumaroyl-CoA ligase; CHS: Chalcone synthase; CHI: Chalcone isomerase; solid red square: strain 135, top naringenin producer in library screening)