TurNuP
Summary¶
Kroll, Rousset, Hu, Liebrand & Lercher (2023) present TurNuP, an organism-independent turnover-number (kcat) prediction model built by the same authors as the DLKcat critique, explicitly designed to address the generalization failures they identified in DLKcat. TurNuP represents whole chemical reactions (not just a single substrate) via differential reaction fingerprints and represents enzymes via a modified, re-trained Transformer protein-sequence model, achieving a test-set R²=0.44 that degrades far more gracefully than DLKcat's as enzyme sequence dissimilarity to the training set increases.
Method¶
- Reaction representation: differential reaction fingerprints (DRFPs) encode the complete reaction (substrates and products), rather than a single substrate structure as in DLKcat.
- Enzyme representation: a Transformer-based protein language model (ESM-1b, further fine-tuned as "ESM-1bESP") encodes the enzyme's amino acid sequence.
- Model: gradient-boosting regression (XGBoost) combining reaction and enzyme representations; the final model averages predictions from an ESM-1bESP-only model and a DRFP-only model.
- Availability: a public web server (no programming required) plus a Python package for batch prediction.
Performance¶
- Test-set accuracy: the full TurNuP model (mean of ESM-1bESP and DRFP sub-models) achieves R²=0.44, MSE=0.81, Pearson r=0.67 on held-out test data — the best of several architecture variants tested (structural reaction fingerprints alone: R²=0.34; ESM-1b alone: R²=0.36; DRFP alone: R²=0.40).
- Generalization by sequence identity: R² is 0.67 for enzymes with 99–100% maximal training-set sequence identity, degrading to R²=0.33 for enzymes below 40% identity — a much smaller collapse than DLKcat's negative R² below 60% identity.
- Generalization by reaction novelty: for reactions absent from the training set entirely, R²=0.35 (versus R²=0.57 for reactions present in training); performance remains moderate (R²=0.20) even for reactions with low similarity (0.4–0.8) to any training reaction, dropping to R²=0.06 only for the 17 test points with near-zero reaction similarity to training data.
- Applied metabolic modeling: parameterizing enzyme-constrained genome-scale metabolic models for multiple yeast species with TurNuP-predicted kcat values improved proteome allocation predictions relative to prior approaches.
Availability¶
Web server: turnup.cs.hhu.de (accepts SMILES, KEGG Compound IDs, or InChI for substrates/products; also reports maximal training-set sequence/reaction similarity as a per-prediction confidence signal). Code: github.com/AlexanderKroll/Kcat_prediction. Data: Zenodo 10.5281/zenodo.7849347.
See Also¶
- DLKcat — predecessor model whose generalization failures TurNuP was designed to address.
- DLKcat Generalization Failure and Train/Test Leakage — the critique motivating TurNuP's evaluation methodology.
- CatPred — reports comparable out-of-distribution kcat accuracy (R²=0.40) using a different architecture.
- CataPro — reproduces TurNuP's benchmark (PCC=0.672 vs. TurNuP's 0.67) when retrained on the same multi-substrate dataset.
- Enzyme Kinetic Parameter Prediction — field overview.