Skip to content

EnzyACT

Summary

EnzyACT predicts whether a single- or multiple-point mutation increases, decreases, or has no effect on enzyme activity, by combining a Graph Convolutional Network over the mutation's local structural neighborhood with ProtT5 sequence embeddings. It was benchmarked against deep-mutational-scanning and literature-curated mutant datasets. Its mechanistic value was further illustrated on cytochrome P450 (CYP2C9), where predicted activity changes were interpreted via molecular dynamics simulation.[1]

Data and Method

  • Training data: 10,998 single-mutation entries (1,303 proteins) curated from the D3DistalMutation database, balanced via a forward/inverse-mutation augmentation scheme (the "S10998" set); a separate multi-mutation training set of 233 proteins ("M576") was built from BRENDA with sequence-identity-based train/test splitting (<32% identity) to limit leakage.[1]
  • Single-point model: a GCN (Deep Graph Library) whose node features are ProtT5-XL-UniRef50 embeddings (1,024-dim per variant; 2,048-dim after concatenating wild-type and mutant) and whose spatial adjacency matrix connects residues within a local radius of the mutation site (reported as 10 Å in Methods, 12 Å in the corresponding figure legend — both values appear in the source and are noted here as an internal inconsistency in the primary paper rather than resolved by this note).[1]
  • Multi-point model: a Support Vector Classification model over the single-point model's outputs plus complementary features, chosen deliberately over a deeper network to avoid overfitting the sparser combinatorial multi-mutant data.[1]
  • Mechanistic case study: wild-type and mutant P450 (CYP2C9) structures were generated via AlphaFold2/PyMOL mutagenesis and subjected to 0.5 µs all-atom MD (NAMD, Amber ff19SB), with RMSD, MM-GBSA/MM-PBSA folding free energy, and dynamic cross-correlation analysis used to interpret predicted activity changes.[1]

Performance

EnzyACT was evaluated on two blind single-mutation test sets — S2814 (2,814 mutations, 420 proteins) and a deep-mutational-scanning set for CYP2C9 (P450, 3,289 mutations from Amorosi et al.) — and a blind multi-mutation set (M167, 167 entries, 95 proteins, <32% identity to training). The paper reports EnzyACT consistently outperforming existing state-of-the-art mutation-effect predictors (including SCANEER, ECNet, and MutCompute) across these independent benchmarks, and correctly identifies known binding and catalytic sites in the P450 case study.[1]

Availability

Data are provided in the paper's Supporting Information; no standalone code repository was identified in this pass.

See Also

  • DETANGO — a complementary approach that further decomposes mutation effects into stability vs. function components rather than predicting a single activity-change label.
  • EZSpecificity — addresses substrate-specificity prediction rather than mutation-driven activity change.
  • Enzyme Kinetic Parameter Prediction — field overview.

Citations

[1] Li, Zhang, Dai, Fan (2024), "EnzyACT: A Novel Deep Learning Method to Predict the Impacts of Single and Multiple Mutations on Enzyme Activity", Journal of Chemical Information and Modeling. Supports: data curation, method, and all performance/case-study claims above. Location: Sections 2 (Materials and Methods), 3 (Results).