Accelerated discovery via a whole-cell model

Abstract

To test the promise of whole-cell modeling to facilitate scientific inquiry, we compared growth rates simulated in a whole-cell model with experimental measurements for all viable single-gene disruption Mycoplasma genitalium strains. Discrepancies between simulations and experiments led to predictions about kinetic parameters of specific enzymes that we subsequently validated. These findings represent, to our knowledge, the first application of whole-cell modeling to accelerate biological discovery.

Figure 1

Model-driven discovery and the quantitative prediction of growth phenotypes. (a) Schematic of a model-driven discovery pipeline as facilitated by a whole-cell model. (b) Simulated (red, n = 5) and experimentally observed (blue, n = 6, technical and biological replicates) specific growth rates (μ) for 86 non-essential gene disruption strains of M. genitalium. Means ± SD are shown, and the absolute value of the difference between model and experiment is shown below on a separate axis. Eighteen genes exhibited significant (heteroskedastic two-tailed t-test and Wilcoxon rank sum test with P ≤ 0.01, listed in Supplemental Table 2) model-experiment discrepancies (top); four of these were small in magnitude (gray). The “lethal zone” indicates the five extremely slow-growing strains which the model called as non-viable. (c) A chromosome map with comparison between model predictions and experimental observations for all 525 of the M. genitalium genes.

Figure 2

The whole-cell model quantitatively predicts rate constants of metabolic reactions. (a,c) Reduced cost analysis of the metabolic fluxes in the ΔthyA (a) and ΔdeoD (c) single-gene disruption strains. Reduced costs for all of the metabolic fluxes in the model are shown, but only the notable costs are labeled. (b,d) Schematic of metabolic reactions which can compensate for those catalyzed by ThyA (b) and DeoD (d). (e) Plot indicating that constraining the flux of the GlpK reaction reduces the ΔMG_039 specific growth rate (orange) more dramatically than in the wild-type (black). (f) Reaction schematic including the MG_039 and GlpK-catalyzed reactions. (g-i) The magnitude of error between the mean model prediction (n = 6) and the mean experimental measurement (n = 5) of Tdk (g), Pdp (h), and GlpK (i) specific growth rates (μ) changes with the kinetic rates of the reactions. The cutoff for acceptable error (dashed horizontal line) for all strains was constrained by the local minimum observed in the ΔMG_039 strain. The model-predicted ranges are indicated by colored horizontal bars just below the x-axis. (j-l) The mapping of Tdk (j), Pdp (k), and GlpK (l) k_cats (at left) to model-predicted specific growth rates of ΔthyA, ΔdeoD, and ΔMG_039 (at right). The colored bars are the same k_cat ranges shown in (g-i), and the colored region indicates the range of simulated specific growth rates determined by the k_cat range. A normal fit to the experimental specific growth rate data is shown at right for comparison to simulated values, and the original estimate of the k_cat used to train the model, together with its corresponding simulated specific growth rate, is shown in red.

Figure 3

Experimental validation of model predictions, and model-driven discovery. (a-c) Hanes-Woolf plots of kinetic assays to measure the v_max of Tdk (a), Pdp (b), and GlpK (c). Error bars represent mean ± SD of three technical replicates per substrate concentration. Linear regression was used to obtain a v_max; k_cat (99% confidence interval indicated on plot) was calculated from the enzyme concentration and v_max. (d) Comparison of k_cat values used to train the model (“Original k_cats” which were estimated from other organisms and not previously measured in M. genitalium), with novel model-based predictions (Fig. 2g–i) and subsequent experimental measurements (Fig. 3a–c). (e) Predicted and measured k_cats were input into the whole-cell model (n = 6) and compared to the experimentally measured specific growth rates and model predictions with original k_cats. P-values were determined by two-tailed t-test, P ≤ 0.01.