EnzymeTuning
Summary¶
Wang, Wang, Zhuang, Wang & Lu (2026, in press at time of ingestion) present EnzymeTuning, a generative adversarial network (GAN)-based "top-down" framework that globally optimizes the kcat values within an enzyme-constrained genome-scale metabolic model (ecGEM) against experimental proteomics and growth-phenotype constraints — in contrast to "bottom-up" sequence-based predictors such as DLKcat and TurNuP, which generate point-estimate kcat values independent of any particular model's system-level behavior. EnzymeTuning also infers protein synthesis rates from literature-derived degradation constants to extend proteome-abundance prediction to previously "dark" (computationally silent) proteins.
Method¶
- Problem framing: sequence-based ("bottom-up") kcat predictors like DLKcat, TurNuP, UniKP, and CatPred/DEKP generate physicochemically-plausible point estimates but neglect system-level constraints (flux balance, proteome capacity) that govern realized in vivo enzyme behavior — directly integrating their outputs into an ecGEM often yields suboptimal phenotype and protein-abundance predictions.
- GAN-based refinement: a generator proposes kcat parameter sets for the ecGEM; a discriminator is trained to distinguish GAN-optimized protein-abundance-prediction distributions from real experimental proteomics distributions, iteratively refining the kcat distribution toward one that is both physiologically plausible and consistent with observed data.
- Protein synthesis rate (v_syn) estimation: literature-reported protein degradation constants (k_deg) are converted into corresponding synthesis-rate bounds via protein mass-conservation relationships, then used as additional ecGEM constraints for a curated subset (empirically, the top ~100) of high-abundance "dark" proteins — over-constraining with all available v_syn values degraded rather than improved performance.
Performance¶
- Benchmarked against Bayesian ecGEM, PRESTO, and CAMEL in S. cerevisiae under carbon-limited chemostat conditions: EnzymeTuning achieved the lowest phenotype-prediction RMSE (0.789, vs. 1.170 for PRESTO and 1.495 for CAMEL) and the highest Pearson correlation between predicted and measured intracellular protein abundances.
- Pathway-level accuracy: strong protein-abundance correlation in central carbon metabolism — pentose phosphate pathway (PCC=0.68, P<0.05), glycolysis/gluconeogenesis (PCC=0.54, P<0.05), pyruvate metabolism (PCC=0.52, P<0.05).
- Proteome coverage: comparable enzyme coverage to the Bayesian method (265 vs. 262 enzymes with predicted abundance), with fewer "extreme" (biologically implausible) enzyme values (46).
- v_syn-constrained extension: adding synthesis-rate constraints for the top ~100 dark proteins increased the number of predictable enzyme abundances to ~400 with a further reduction in extreme values.
- Generalization: validated across multiple organisms (S. cerevisiae, Kluyveromyces lactis, Kluyveromyces marxianus, Yarrowia lipolytica, E. coli) and multiple nitrogen sources (glutamate, iso-leucine, phenylalanine), with consistent improvement patterns.
- Computational cost: ~24 hours runtime, comparable to the Bayesian method but substantially higher than PRESTO (~7h) or CAMEL (~6h); the authors note the GAN training is parallelizable on HPC infrastructure.
Availability¶
Code, data, and growth-data supplement: github.com/hongzhonglu/EnzymeTuning.
See Also¶
- DLKcat and TurNuP — "bottom-up" sequence-based kcat predictors whose point estimates EnzymeTuning's top-down GAN approach is designed to refine against system-level, in vivo constraints.
- Enzyme Kinetic Parameter Prediction — field overview distinguishing bottom-up sequence prediction from top-down model-based calibration.