Skip to content

CataPro

Summary

Wang, Xie, Wu, et al. (2025) introduce CataPro alongside a set of "unbiased" benchmark datasets for kcat, Km, and kcat/Km — built by clustering enzymes at 40% sequence identity and partitioning clusters (rather than individual entries) into ten cross-validation folds, so that near-duplicate sequences cannot appear across both train and test partitions, the same leakage failure mode identified in DLKcat by Kroll & Lercher. CataPro combines pretrained protein/molecule embeddings (ProtT5, MolT5) with molecular fingerprints (RDKit, MACCS) and, in an applied enzyme-mining case study, identified an oxygenase (SsCSO) with 19.53× the activity of the initial candidate enzyme (CSO2) for vanillin biosynthesis, then engineered a further 3.34× improvement via CataPro-guided mutant ranking.

Method

  • Unbiased benchmark construction: enzyme-substrate entries from BRENDA and SABIO-RK, clustered by protein sequence identity (CD-HIT, 0.4 threshold) into ten folds for cross-validation — retraining DLKcat and UniKP under the same unbiased splits for fair comparison.
  • Features: ProtT5 protein embeddings, MolT5 substrate embeddings, and MACCS molecular fingerprints, combined via a feature-wise attention module.
  • kcat/Km prediction: rather than simply dividing predicted kcat by predicted Km (which compounds both models' errors), CataPro pre-trains kcat and Km sub-models then applies a neural-network correction term to the kcat/Km estimate.
  • Model selection: an Extra Trees regressor (consistent with prior work showing tree ensembles outperform neural networks for this feature scale) was used for final kcat/Km training after ablation testing.

Performance

  • Unbiased 10-fold benchmark: CataPro shows clearly improved accuracy and generalization (PCC, SCC, RMSE) relative to retrained DLKcat and UniKP baselines on the same unbiased splits (Fig. 2), and outperforms both on kcat/Km specifically (PCC=0.413 vs. lower for DLKcat/UniKP).
  • Cross-validation methodology matters: when instead evaluated on randomly-split (non-clustered) folds — the evaluation style used by most prior literature — UniKP actually outperformed CataPro on kcat prediction, demonstrating that random splitting can produce overly optimistic, non-generalizable performance estimates for any of these models.
  • Comparison to TurNuP on multi-substrate reactions: on TurNuP's own multi-substrate kcat test set, CataPro (averaging per-substrate predictions) achieved PCC=0.661, R²=0.415 — close to TurNuP's PCC=0.67, R²=0.44; retraining CataPro directly on TurNuP's training data and folds narrowed this further (PCC=0.672, R²=0.451, MSE=0.783 vs. TurNuP's PCC=0.67, R²=0.44, MSE=0.81), and reproduced TurNuP's generalization pattern by sequence identity (R²=0.333 at 0–40% identity, closely matching TurNuP's R²=0.33).
  • Enzyme discovery case study: starting from ferulic-acid-decarboxylase-derived 4-vinylguaiacol (4-VG), CataPro-guided screening of BLAST-retrieved homologs (structurally clustered via TM-align) identified SsCSO (from Sphingobium sp.) with 19.53× the oxidation activity of the initial candidate CSO2; subsequent CataPro-guided mutagenesis of SsCSO yielded a further 3.34× activity increase.
  • Mutation ranking: on deep mutational scanning (DMS) datasets, CataPro's kcat and kcat/Km sub-models approach or exceed the ranking accuracy (Spearman correlation) of Position-Specific Scoring Matrices (PSSM), a strong evolutionary-conservation baseline, outperforming both UniKP and DLKcat on the same DMS panels.

Availability

Code, unbiased benchmark datasets, and documentation: github.com/zchwang/CataPro. Data: Zenodo.

See Also

Citations

[1] Wang, Z., Xie, D., Wu, D., Luo, X., Wang, S., Li, Y., Yang, Y., Li, W., & Zheng, L. (2025). Robust enzyme discovery and engineering with deep learning using CataPro. Nature Communications, 16, 3489.