Machine learning applied to enzyme turnover numbers reveals protein structural correlates and improves metabolic models

Abstract

Knowing the catalytic turnover numbers of enzymes is essential for understanding the growth rate, proteome composition, and physiology of organisms, but experimental data on enzyme turnover numbers is sparse and noisy. Here, we demonstrate that machine learning can successfully predict catalytic turnover numbers in Escherichia coli based on integrated data on enzyme biochemistry, protein structure, and network context. We identify a diverse set of features that are consistently predictive for both in vivo and in vitro enzyme turnover rates, revealing novel protein structural correlates of catalytic turnover. We use our predictions to parameterize two mechanistic genome-scale modelling frameworks for proteome-limited metabolism, leading to significantly higher accuracy in the prediction of quantitative proteome data than previous approaches. The presented machine learning models thus provide a valuable tool for understanding metabolism and the proteome at the genome scale, and elucidate structural, biochemical, and network properties that underlie enzyme kinetics.

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

“…We further found that kapp,max outperforms kcat in vitro in MOMENT and ME models across all growth conditions ( Figure 5). When comparing median-imputed kcat parameterizations to those using supervised machine learning, we found that machine learning reduces RMSE on log10 scale by 38% for kapp,max and 10% for kcat in vitro, confirming the utility of this approach 16 . Figure 5: Performance in mechanistic models of proteome allocation.…”

“…In order to validate the estimated in vivo turnover numbers in a genome-scale model that contains over three thousand direction-specific reactions, we first needed to extrapolate the data to the genome scale. We used supervised machine learning on a diverse enzyme data set 16 that includes data on enzyme network context, enzyme 3D structure, and enzyme biochemistry to achieve this goal. An ensemble model of an elastic net, random forest, and neural network 16 showed good performance in cross-validation for the in vivo turnover numbers, where the highest performance was achieved for kapp,max that was obtained from the 21 KO strains ( Figure 4).…”

“…We used supervised machine learning on a diverse enzyme data set 16 that includes data on enzyme network context, enzyme 3D structure, and enzyme biochemistry to achieve this goal. An ensemble model of an elastic net, random forest, and neural network 16 showed good performance in cross-validation for the in vivo turnover numbers, where the highest performance was achieved for kapp,max that was obtained from the 21 KO strains ( Figure 4). Taking the maximum of kapp,max from this study and that of Davidi et al 3 did not improve model performance, even though it resulted in the largest training set.…”

“…Figure 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 16 (see Methods). Data for kcat in vitro was taken from Heckmann et al 16 .…”

“…Turnover numbers are successfully used in genome-scale metabolic models to constrain metabolic fluxes by a limited cellular protein budget [17][18][19] or the balance of translation and dilution of proteins [20][21][22] . The kapp,max obtained from diverse growth conditions was previously used successfully in genome-scale metabolic models, showing that the performance of protein abundance predictions of models using kapp,max was significantly higher than that of models using in vitro kcats 16 . A major drawback of this analysis lies in the validation of the metabolic model which used data 23 that was also utilized in obtaining kapp,max 3 , posing the risk of circular reasoning through data leakage.…”