Skip to content

EZSpecificity

Summary

EZSpecificity predicts which substrate(s) an enzyme will preferentially act on — a distinct problem from turnover-rate prediction, since millions of characterized enzymes still lack reliable substrate-specificity annotation. It combines ESM-2 sequence embeddings with an SE(3)-equivariant message-passing network and a cross-attention layer over active-site residues and substrate atoms. It was validated experimentally on eight halogenases.[1]

Method

  • Sequence embeddings are generated with ESM-2.[1]
  • Structural/interaction embeddings use a message-passing neural network within an SE(3)-equivariant framework, so the model's assessment of active-site geometry relative to a substrate is invariant to the molecule's absolute translation/rotation.[1]
  • A cross-attention mechanism weights the interaction importance between specific active-site residues and individual substrate atoms.[1]
  • The model is trained on a purpose-built database of enzyme-substrate interactions annotated at both sequence and structural levels.[1]

Performance

In experimental validation across eight halogenases and 78 candidate substrates, EZSpecificity achieved 91.7% accuracy in identifying the single reactive substrate from the decoy pool, compared with 58.3% for the prior state-of-the-art enzyme-substrate specificity model.[1]

Availability

Full text of the primary source was not accessible in this pass (Nature is subscription-gated; only the abstract was inspected via the publisher's page). No code repository was identified. Confidence is set to medium and should be revisited once the full text or a preprint is available — the architecture and headline validation numbers above come directly from the abstract, but the paper's Methods, additional benchmarks, and code/data availability could not be verified.

See Also

  • EnzyACT — a related graph/language-model architecture for mutation-impact prediction on enzyme activity rather than substrate choice.
  • Enzyme Kinetic Parameter Prediction — field overview of related sequence-to-kinetics models.

Citations

[1] Cui, Su, Dean, Yu, Zhang et al. (2025), "Enzyme specificity prediction using cross-attention graph neural networks", Nature. Supports: method summary and the halogenase validation figures above. Location: Abstract only (full text not accessed — publisher paywall).