Heckmann Enzyme Turnover ML Model
Summary¶
Heckmann, Lloyd, Mih, et al. (2018) predict E. coli enzyme turnover numbers — both the traditional in vitro kcat and the in vivo apparent maximal turnover rate, kapp,max — from an integrated feature set spanning enzyme structure (active site geometry, solvent accessibility, structural disorder), network context (metabolic flux, "generalist" multi-reaction usage), and biochemistry (EC number, Km, thermodynamic efficiency, metabolite concentrations). Predating the deep-learning kcat predictors (DLKcat, TurNuP) by several years, this classical-ML ensemble (random forest, elastic net, neural network) achieved substantially higher accuracy for kapp,max than for kcat in vitro, and using its predictions to parameterize proteome-allocation genome-scale metabolic models markedly improved their accuracy against measured proteomics data.
Method¶
- Feature classes: network properties (in silico flux from parsimonious FBA on the E. coli GEM iML1515; "generalist" score for enzymes catalyzing multiple reactions), structural properties (from PDB structures and I-TASSER homology models: global structural disorder, molecular weight, active site depth/solvent exposure/hydrophobicity/secondary structure, from the Catalytic Site Atlas), biochemistry (EC number, thermodynamic efficiency, Km, metabolite concentrations), and assay conditions (pH, temperature — for kcat in vitro models only).
- Training data: 215 complete kcat in vitro observations (497 after feature imputation) from BRENDA, SABIO-RK, and MetaCyc; 133 complete kapp,max observations (234 after imputation), where kapp,max is the maximum apparent turnover rate across diverse growth conditions estimated from proteomics + flux-balance analysis in prior work.
- Models: linear regression, PLSR, elastic net, random forest, and a deep neural network, evaluated via repeated 5-fold cross-validation plus an independent 20% test set; a final ensemble averaged predictions from elastic net, random forest, and neural network (in 12 variants combining different imputation strategies).
Performance¶
- kapp,max is far more predictable than kcat in vitro: mean cross-validated R²=0.76 for kapp,max versus R²=0.31 for kcat in vitro — attributed to kcat in vitro's higher measurement noise from variable assay conditions and incomplete enzyme saturation.
- Feature importance: in silico flux is the single most important predictor for both kcat in vitro and kapp,max, consistent with the hypothesis that evolutionary selection pressure (proportional to how much an enzyme is actually used) shapes catalytic turnover; the "generalist" property (multi-reaction enzymes have lower turnover) and active-site structural features (depth, solvent accessibility) were also consistently important in both models. Km was strongly predictive for kcat in vitro but not for kapp,max.
- Downstream metabolic model improvement: parameterizing two mechanistic proteome-allocation GEM frameworks (MOMENT and an ME model) with kapp,max-based (rather than kcat in vitro-based) parameters reduced proteomics-prediction error by 43% in both frameworks; further replacing median-imputed kapp,max with the ML ensemble's predicted kapp,max reduced RMSE by an additional 34% (MOMENT) and 20% (ME model). The equivalent gain from ML-predicted kcat in vitro was much smaller (7% and 1% RMSE reduction respectively), consistent with its lower cross-validated accuracy.
Significance¶
This 2018 study is an early precedent for the sequence/structure-to-kcat machine learning paradigm later scaled up by deep learning models such as DLKcat, TurNuP, and CatPred — notably using classical ML with interpretable structural features rather than end-to-end deep representations, and explicitly validating its value via downstream proteome-allocation model accuracy rather than only kcat prediction accuracy in isolation. Structural feature extraction here (PDB structures, I-TASSER homology models, Catalytic Site Atlas annotations) closely parallels the GEM-PRO methodology used in Recon3D and the Multi-scale Erythrocyte Metabolism Platform from the same research group.
See Also¶
- Enzyme Kinetic Parameter Prediction — field overview; this model is an early classical-ML precedent to the deep learning tools reviewed there.
- DLKcat and TurNuP — later deep-learning kcat predictors.
- Recon3D and Multi-scale Erythrocyte Metabolism Platform — related genome-scale metabolic modeling work from the same lab, using the same protein-structural-integration philosophy (GEM-PRO).