Skip to content

KcatNet

Summary

KcatNet predicts enzyme turnover numbers (kcat) at genome scale from paired enzyme-sequence and substrate representations, using a geometric deep learning design intended to localize the catalytic pocket rather than treat the enzyme as an undifferentiated sequence. It was applied to reconstruct enzyme-constrained genome-scale metabolic models (ecGEMs) across yeast species. It was also used to guide directed evolution of α-glucosidase.[1]

Architecture

KcatNet has three components:

  1. Enzyme-substrate encoder: enzyme residue-level features from ProtT5 and ESM2 (two pretrained protein language models); substrates encoded at both atomic level (atom type/property embeddings) and molecular level (a pretrained SMILES transformer).[1]
  2. Graph-based residue-level partition module: the enzyme structure (or, when unresolved, an ESM2-derived contact map) is modeled as a graph and processed with a graph convolutional network, then partitioned into spatially proximal residue clusters via graph partitioning, capturing the local geometry of candidate catalytic pockets.[1]
  3. Interaction module: an attention-based MLP that iteratively refines substrate and enzyme-partition embeddings against each other to predict kcat, while also yielding a per-residue importance/activation score.[1]

Findings

  • Interpretability: activation scores correctly highlighted catalytic residues for test enzymes (e.g. an enolase from Pelagibaca bermudensis, PDB 4H2H) and separately flagged substrate-binding (non-catalytic) residues for others (e.g. E. coli dTDP-glucose 4,6-dehydratase, PDB 1BXK), consistent with UniProt binding/active-site annotations.[1]
  • Genome-scale metabolic modeling: integrating KcatNet-predicted kcat values into enzyme-constrained GEMs (ecGEMs) for yeast/fungal species improved growth-rate predictions relative to the original ecGEM pipeline (which achieved kcat coverage for only ~40% of enzymes vs. ~80% with deep-learning predictions) and, compared with ecGEMs parameterized by DLKcat, reduced growth-rate mean squared error in 16 of 22 environment-species combinations.[1]
  • Directed evolution (α-glucosidase): of nine single-point mutations KcatNet predicted would enhance catalytic efficiency, seven were experimentally validated to exceed wild-type kcat.[1]

Availability

This is a preprint, not yet peer-reviewed; no public code repository was identified in this pass, so confidence is set to medium.

See Also

  • DLKcat — the ecGEM baseline KcatNet is directly compared against for proteome-allocation/growth-rate accuracy.
  • RealKcat — a contemporary kcat/Km predictor targeting mutation sensitivity via gradient boosting rather than genome-wide structural generalization.
  • Structural Km Predictor — a related graph-neural-network approach for the complementary Km parameter.
  • Enzyme Kinetic Parameter Prediction — field overview.

Citations

[1] Pan, Cui, Koh, Bi, Wang et al. (2025), "KcatNet: A Geometric Deep Learning Framework for Genome-Wide Prediction of Enzyme Catalytic Efficiency", bioRxiv Supports: architecture, interpretability case studies, ecGEM and directed-evolution results above. Location: Abstract; Results (model overview; catalytic-pocket interpretability; genome-scale ecGEM application; α-glucosidase directed evolution).