A novel benchmark dataset for enzyme function prediction reveals the limitations of state-of-the-art models

In plain English

The authors mutate active sites while keeping much of an enzyme sequence recognizable, then ask whether annotation models incorrectly assign the original Enzyme Commission label.

What they found

• DIAMOND and CLEAN exceeded 90% false-positive rates for low-perturbation decoys.
• DeepEC improved at larger perturbations, but all tested approaches struggled with targeted active-site disruption.

The catch

• This is a preprint and has not been peer reviewed.
• The decoys are expected to be inactive computationally; they were not experimentally validated as non-functional.

Methods at a glance

Study type

Computational benchmark preprint

Biological system

Enzyme catalytic sites

Organism

Multiple experimentally annotated enzyme sources

Dataset

EnzymARC structure-guided decoy benchmark

Sample size

79,505 annotated entries used by the project pipeline

Model architecture

Benchmark; evaluated DIAMOND, CLEAN, and DeepEC

Training objective

Evaluate resistance to active-site-disruption decoys

Code

Benchmark generation and evaluation code are available from the linked GitHub repository.

Data

The project repository documents the EnzymARC benchmark files and generation process.

Tasks and metrics

Tasks: Enzyme Commission number prediction on putatively non-functional decoys

Metrics: False-positive rate

Baselines: DIAMOND; CLEAN; DeepEC

Principal findings

• DIAMOND and CLEAN had false-positive rates above 90% on low-perturbation decoys.

What remains unresolved

External validation: Not reported

• The work is a preprint.
• Putative inactivity is not confirmed by biochemical experiments.