Kinetic profiling of metabolic specialists demonstrates stability and consistency of in vivo enzyme turnover numbers

Abstract

Enzyme turnover numbers (k_cats) are essential for a quantitative understanding of cells. Because k_cats are traditionally measured in low-throughput assays, they can be inconsistent, labor-intensive to obtain, and can miss in vivo effects. We use a data-driven approach to estimate in vivo k_cats using metabolic specialist Escherichia coli strains that resulted from gene knockouts in central metabolism followed by metabolic optimization via laboratory evolution. By combining absolute proteomics with fluxomics data, we find that in vivo k_cats are robust against genetic perturbations, suggesting that metabolic adaptation to gene loss is mostly achieved through other mechanisms, like gene-regulatory changes. Combining machine learning and genome-scale metabolic models, we show that the obtained in vivo k_cats predict unseen proteomics data with much higher precision than in vitro k_cats. The results demonstrate that in vivo k_cats can solve the problem of inconsistent and low-coverage parameterizations of genome-scale cellular models.

Keywords: gene knockout; in vivo; kcat; proteomics; turnover number.

Fig. 1.

Approach for obtaining k_cat in vivo from metabolic specialists: KO of enzymes in central metabolism was followed by ALE to obtain 21 strains that had diverse flux profiles, while achieving high growth rates (–9). Fluxomics and proteomics data were then integrated for the evolved strains to obtain the maximum k_app across the 21 strains (k_app,max) for each enzyme that could be mapped uniquely. The obtained k_app,max vector was then extrapolated to genome scale via supervised machine learning and used to parameterize genome-scale metabolic models. The resulting genome-scale models were then validated on unseen proteomics data.

Fig. 2.

Apparent catalytic rates cluster by genetic background and exhibit diversity across strains. (A) Data on k_app in each of the 21 strains projected onto the first two principal components. Only reaction−strain combinations for which k_app was available in all strains were used, resulting in 214 reactions used in the analysis. Data were centered and scaled before conducting principal component analysis. (B) Distribution of ranges of k_app across reactions. The log₂ of the ratio between the highest and the lowest k_app per reaction is shown.

Fig. 3.

Estimates of in vivo turnover numbers are consistent. (A) Comparison between k_app,max obtained from KO strains (this study) and k_app,max from growth conditions (3). MAE, mean absolute error. (B) Number of reactions for which k_app,max was obtained in KO strains (this study) and varying growth conditions (3). (C) Comparison between k_app,max obtained from KO strains and in vitro k_cats. (D) Comparison between k_app,max obtained from KO strains and in vitro k_cats (3). Horizontal lines are 95% CIs determined by 500 parametric bootstrap samples (see Materials and Methods). Points are marked red when the compared value falls into the 95% CI of k_app,max from KO strains and are colored blue if the compared value does not fall into the CI. Points are labeled with reaction IDs as given in the iJO1366 reconstruction (51) if the values differ by more than one order of magnitude. Data on k_cat in vitro shown in C and D were taken from Davidi et al. (3) to allow for comparison between the studies. Davidi et al. (3) obtained this in vitro dataset from the Braunschweig Enzyme Database (BRENDA) (52) and utilized the maximal k_cat in cases where multiple sources were available for the same enzyme.

Fig. 4.

Performance of machine learning models on different sources of turnover numbers. The performance is estimated in five-times repeated five-fold cross-validation in elastic net, random forest, and a neural network (15) (see Materials and Methods). Data for k_cat in vitro were taken from Heckmann et al. (15).

Fig. 5.

Performance of turnover number vectors in mechanistic models of proteome allocation. The MOMENT algorithm and an ME model were parameterized with different sources of turnover numbers [including Davidi et al. (3)]. Growth on different carbon sources was simulated with the two algorithms to predict relative protein weight fractions of metabolic enzymes. The protein abundance predictions were then compared to proteomics data in the respective growth condition published by Schmidt et al. (22) using the RMSE on log₁₀ scale. Machine learning model predictions were based on the protocol introduced by Heckmann et al. (15), but all models were newly trained on the data produced in this study (Dataset S1).